diff --git a/.env.example b/.env.example
index b5ec401c..8bc92cb9 100644
--- a/.env.example
+++ b/.env.example
@@ -15,6 +15,14 @@ BRAINY_WORKER_POLL_INTERVAL=2s
 #
 # Same-pin Mem0 Platform evals (evals/run_competitor_benchmark.py, --systems mem0):
 # MEM0_API_KEY=[REDACTED]
+#
+# OpMem external lanes (evals/run_opmem_external_live.py; see docs/research/opmem/external-lanes-setup.md):
+# OPENAI_API_KEY=          # mem0-oss, letta, langmem (pinned gpt-4.1-mini + text-embedding-3-small)
+# ZEP_API_KEY=             # Zep Cloud free tier only ($0 paid ceiling)
+# LETTA_BASE_URL=http://127.0.0.1:8283
+# LETTA_APP_SERVER_TOKEN=  # optional for local letta server --no-secure
+# DATABASE_URL=postgresql://opmem:opmem@127.0.0.1:5432/opmem_langmem
+# OPMEM_QDRANT_ROOT=/tmp/opmem-qdrant
 
 # HTTP boundary (API). Defaults: 5 MiB body, 10s header / 30s read / provider+60s write / 120s idle.
 # BRAINY_MAX_BODY_BYTES=5242880
diff --git a/docs/benchmarks/artifacts/opmem-external-lanes/PHASE2_SUMMARY.md b/docs/benchmarks/artifacts/opmem-external-lanes/PHASE2_SUMMARY.md
new file mode 100644
index 00000000..a6815226
--- /dev/null
+++ b/docs/benchmarks/artifacts/opmem-external-lanes/PHASE2_SUMMARY.md
@@ -0,0 +1,26 @@
+# OpMem external lanes — execution summary
+
+**Protocol:** 13 tasks / 47 steps (`fixtures/opmem/`).
+
+## Full comparison
+
+| Lane | Run ID | Score | Infra errors | SDK / infra |
+| --- | --- | ---: | ---: | --- |
+| mem0-oss | `opmem-phase2-full-mem0` | **9/13** | 0 | `mem0ai` 2.1.0, local Qdrant path |
+| langmem | `opmem-phase2-full1` | **8/13** | 0 | Postgres + pgvector |
+| letta | `opmem-phase2-letta-full` | **5/13** | 0 | `letta==0.11.7` |
+| zep | `opmem-phase2-zep-full` | **3/9\*** | 0 | `zep-cloud==3.28.0`, free tier, $0 paid |
+
+**\* Zep (publication line):** Score **3/9** counts only tasks where forget is in scope for the lane. **Not supported** (listed separately, not counted as failures): `iso03_forget_isolated`, `sup01_basic_forget`, `sup02_targeted_forget`, `sup03_durable_forget` — the `zep-cloud` 3.28 thread/graph adapter has no delete path for OpMem `forget`. **Raw harness** on the unchanged artifact `live-full-opmem-phase2-zep-full.json` remains **3/13**. See `zep-score-interpretation-opmem-phase2-zep-full.json`.
+
+### Zep supported-task breakdown (`opmem-phase2-zep-full`)
+
+| Status | Tasks |
+| --- | --- |
+| **Passed (3)** | `cor02_correction_stickiness`, `cor03_revised_retrievable`, `dup01_idempotent_remember` |
+| **Failed, scored (6)** | `cor01_basic_revision`, `iso01_subject_isolation`, `iso02_tenant_isolation`, `upd01_stale_fact`, `upd02_preference_change`, `upd03_state_supersession` |
+| **Not supported — forget (4)** | `iso03_forget_isolated`, `sup01_basic_forget`, `sup02_targeted_forget`, `sup03_durable_forget` |
+
+## Spend
+
+Payload worst-case OpenAI ~$0.15 vs $3 ceiling (`cost-worksheet-payload-budget.json`). Zep lane $0 paid.
diff --git a/docs/benchmarks/artifacts/opmem-external-lanes/commands.log b/docs/benchmarks/artifacts/opmem-external-lanes/commands.log
new file mode 100644
index 00000000..08cf05a7
--- /dev/null
+++ b/docs/benchmarks/artifacts/opmem-external-lanes/commands.log
@@ -0,0 +1,39 @@
+
+# 2026-09-19T08:28:55Z run_id=opmem-phase2-smoke1
+r u n _ o p m e m _ e x t e r n a l _ l i v e . p y   - - s m o k e   - - l a n e s   m e m 0 - o s s , l a n g m e m , z e p , l e t t a   - - r u n - i d   o p m e m - p h a s e 2 - s m o k e 1
+
+# 2026-09-19T08:29:18Z run_id=opmem-phase2-smoke2
+r u n _ o p m e m _ e x t e r n a l _ l i v e . p y   - - s m o k e   - - l a n e s   m e m 0 - o s s , l a n g m e m   - - r u n - i d   o p m e m - p h a s e 2 - s m o k e 2
+
+# 2026-09-19T08:30:32Z run_id=opmem-phase2-full1
+r u n _ o p m e m _ e x t e r n a l _ l i v e . p y   - - f u l l   - - l a n e s   m e m 0 - o s s , l a n g m e m   - - r u n - i d   o p m e m - p h a s e 2 - f u l l 1   - - s k i p - p r e f l i g h t
+
+# 2026-09-19T08:31:22Z run_id=opmem-phase2-full-mem0
+r u n _ o p m e m _ e x t e r n a l _ l i v e . p y   - - f u l l   - - l a n e s   m e m 0 - o s s   - - r u n - i d   o p m e m - p h a s e 2 - f u l l - m e m 0   - - s k i p - p r e f l i g h t
+
+# 2026-09-21T17:43:52Z run_id=opmem-phase2-letta-smoke
+r u n _ o p m e m _ e x t e r n a l _ l i v e . p y   - - s m o k e   - - l a n e s   l e t t a   - - r u n - i d   o p m e m - p h a s e 2 - l e t t a - s m o k e
+
+# 2026-09-21T17:43:56Z run_id=opmem-phase2-zep-smoke
+r u n _ o p m e m _ e x t e r n a l _ l i v e . p y   - - s m o k e   - - l a n e s   z e p   - - r u n - i d   o p m e m - p h a s e 2 - z e p - s m o k e
+
+# 2026-09-21T17:44:03Z run_id=opmem-phase2-letta-full
+r u n _ o p m e m _ e x t e r n a l _ l i v e . p y   - - f u l l   - - l a n e s   l e t t a   - - r u n - i d   o p m e m - p h a s e 2 - l e t t a - f u l l
+
+# 2026-09-23T17:09:16Z run_id=opmem-phase2-zep-smoke-20260923
+r u n _ o p m e m _ e x t e r n a l _ l i v e . p y   - - s m o k e   - - l a n e s   z e p   - - r u n - i d   o p m e m - p h a s e 2 - z e p - s m o k e - 2 0 2 6 0 9 2 3
+
+# 2026-09-23T17:09:18Z run_id=opmem-phase2-zep-full
+r u n _ o p m e m _ e x t e r n a l _ l i v e . p y   - - f u l l   - - l a n e s   z e p   - - r u n - i d   o p m e m - p h a s e 2 - z e p - f u l l
+
+# 2026-09-23T17:55:34Z run_id=opmem-phase2-zep-smoke-20260923b
+r u n _ o p m e m _ e x t e r n a l _ l i v e . p y   - - s m o k e   - - l a n e s   z e p   - - r u n - i d   o p m e m - p h a s e 2 - z e p - s m o k e - 2 0 2 6 0 9 2 3 b
+
+# 2026-09-23T17:56:27Z run_id=opmem-phase2-zep-smoke-20260923c
+r u n _ o p m e m _ e x t e r n a l _ l i v e . p y   - - s m o k e   - - l a n e s   z e p   - - r u n - i d   o p m e m - p h a s e 2 - z e p - s m o k e - 2 0 2 6 0 9 2 3 c
+
+# 2026-09-23T17:56:54Z run_id=opmem-phase2-zep-smoke-final
+r u n _ o p m e m _ e x t e r n a l _ l i v e . p y   - - s m o k e   - - l a n e s   z e p   - - r u n - i d   o p m e m - p h a s e 2 - z e p - s m o k e - f i n a l
+
+# 2026-09-23T17:58:04Z run_id=opmem-phase2-zep-full
+r u n _ o p m e m _ e x t e r n a l _ l i v e . p y   - - f u l l   - - l a n e s   z e p   - - r u n - i d   o p m e m - p h a s e 2 - z e p - f u l l
diff --git a/docs/benchmarks/artifacts/opmem-external-lanes/cost-worksheet-dry-run.json b/docs/benchmarks/artifacts/opmem-external-lanes/cost-worksheet-dry-run.json
new file mode 100644
index 00000000..c16ab779
--- /dev/null
+++ b/docs/benchmarks/artifacts/opmem-external-lanes/cost-worksheet-dry-run.json
@@ -0,0 +1,151 @@
+{
+  "generated_utc": "2026-09-19T08:29:00Z",
+  "task_count": 13,
+  "step_count": 47,
+  "rates": {
+    "gpt-4.1-mini-2025-04-14": {
+      "input_per_m": 0.4,
+      "output_per_m": 1.6
+    },
+    "text-embedding-3-small": {
+      "per_m": 0.02
+    }
+  },
+  "pinned_models": {
+    "chat": "gpt-4.1-mini-2025-04-14",
+    "embed": "text-embedding-3-small"
+  },
+  "ceilings_usd": {
+    "mem0-oss": 1.0,
+    "letta": 1.0,
+    "langmem": 1.0,
+    "zep": 0.0,
+    "total": 3.0
+  },
+  "lanes": [
+    {
+      "lane": "langmem",
+      "operations": {
+        "remember": 21,
+        "recall": 19,
+        "revise": 3,
+        "forget": 4,
+        "reset": 13,
+        "http_calls": 47
+      },
+      "tokens": {
+        "chat_model": "gpt-4.1-mini-2025-04-14",
+        "embed_model": "text-embedding-3-small",
+        "chat_input_tokens": 409,
+        "chat_output_tokens": 66,
+        "embed_tokens": 191,
+        "estimated_usd": 0.000273
+      },
+      "product_constraints": [],
+      "credentials_present": false,
+      "missing_credentials": [
+        "DATABASE_URL"
+      ],
+      "projected_usd": 0.000273,
+      "ceiling_usd": 1.0,
+      "within_ceiling": true,
+      "live_run_allowed": false,
+      "live_run_note": "Phase 1 dry-run only; enable after credentials present and measured estimate confirmed under ceiling."
+    },
+    {
+      "lane": "letta",
+      "operations": {
+        "remember": 21,
+        "recall": 19,
+        "revise": 3,
+        "forget": 4,
+        "reset": 13,
+        "http_calls": 47
+      },
+      "tokens": {
+        "chat_model": "gpt-4.1-mini-2025-04-14",
+        "embed_model": "text-embedding-3-small",
+        "chat_input_tokens": 409,
+        "chat_output_tokens": 66,
+        "embed_tokens": 191,
+        "estimated_usd": 0.000273
+      },
+      "product_constraints": [],
+      "credentials_present": false,
+      "missing_credentials": [
+        "LETTA_APP_SERVER_TOKEN"
+      ],
+      "projected_usd": 0.000273,
+      "ceiling_usd": 1.0,
+      "within_ceiling": true,
+      "live_run_allowed": false,
+      "live_run_note": "Phase 1 dry-run only; enable after credentials present and measured estimate confirmed under ceiling."
+    },
+    {
+      "lane": "mem0-oss",
+      "operations": {
+        "remember": 21,
+        "recall": 19,
+        "revise": 3,
+        "forget": 4,
+        "reset": 13,
+        "http_calls": 47
+      },
+      "tokens": {
+        "chat_model": "gpt-4.1-mini-2025-04-14",
+        "embed_model": "text-embedding-3-small",
+        "chat_input_tokens": 409,
+        "chat_output_tokens": 66,
+        "embed_tokens": 191,
+        "estimated_usd": 0.000273
+      },
+      "product_constraints": [],
+      "credentials_present": true,
+      "missing_credentials": [],
+      "projected_usd": 0.000273,
+      "ceiling_usd": 1.0,
+      "within_ceiling": true,
+      "live_run_allowed": false,
+      "live_run_note": "Phase 1 dry-run only; enable after credentials present and measured estimate confirmed under ceiling."
+    },
+    {
+      "lane": "zep",
+      "operations": {
+        "remember": 21,
+        "recall": 19,
+        "revise": 3,
+        "forget": 4,
+        "reset": 13,
+        "http_calls": 47
+      },
+      "tokens": {
+        "chat_model": "gpt-4.1-mini-2025-04-14",
+        "embed_model": "text-embedding-3-small",
+        "chat_input_tokens": 0,
+        "chat_output_tokens": 0,
+        "embed_tokens": 0,
+        "estimated_usd": 0.0
+      },
+      "product_constraints": [
+        "lane zep cannot pin chat model gpt-4.1-mini-2025-04-14",
+        "lane zep cannot pin embed model text-embedding-3-small",
+        "Zep Cloud memory path does not accept OpenAI chat/embed model pins; stop before substituting \u2014 record scores only under Zep-native models."
+      ],
+      "credentials_present": false,
+      "missing_credentials": [
+        "ZEP_API_KEY"
+      ],
+      "projected_usd": 0.0,
+      "ceiling_usd": 0.0,
+      "within_ceiling": true,
+      "live_run_allowed": false,
+      "live_run_note": "Phase 1 dry-run only; enable after credentials present and measured estimate confirmed under ceiling."
+    },
+    {
+      "lane": "_total_openai_estimate",
+      "projected_usd": 0.000819,
+      "ceiling_usd": 3.0,
+      "within_ceiling": true
+    }
+  ]
+}
diff --git a/docs/benchmarks/artifacts/opmem-external-lanes/cost-worksheet-payload-budget.json b/docs/benchmarks/artifacts/opmem-external-lanes/cost-worksheet-payload-budget.json
new file mode 100644
index 00000000..b61ad6cc
--- /dev/null
+++ b/docs/benchmarks/artifacts/opmem-external-lanes/cost-worksheet-payload-budget.json
@@ -0,0 +1,87 @@
+{
+  "mode": "payload_budget",
+  "task_count": 13,
+  "step_count": 47,
+  "pinned_chat_model": "gpt-4.1-mini-2025-04-14",
+  "pinned_embed_model": "text-embedding-3-small",
+  "rate_source_url": "https://openai.com/api/pricing/",
+  "rate_source_note": "gpt-4.1-mini-2025-04-14: $0.40/1M input, $1.60/1M output; text-embedding-3-small: $0.02/1M tokens",
+  "ceilings_usd": {
+    "mem0-oss": 1.0,
+    "letta": 1.0,
+    "langmem": 1.0,
+    "zep": 0.0,
+    "total": 3.0
+  },
+  "lanes": [
+    {
+      "lane": "langmem",
+      "worst_case_usd": 0.017537,
+      "token_totals": {
+        "chat_input_tokens": 1447,
+        "chat_output_tokens": 9600,
+        "embed_tokens": 203
+      },
+      "line_count": 47,
+      "product_constraints": [],
+      "retry_margin": 1.1,
+      "projected_worst_case_usd": 0.017537,
+      "ceiling_usd": 1.0,
+      "within_ceiling": true
+    },
+    {
+      "lane": "letta",
+      "worst_case_usd": 0.013526,
+      "token_totals": {
+        "chat_input_tokens": 1930,
+        "chat_output_tokens": 7200,
+        "embed_tokens": 203
+      },
+      "line_count": 47,
+      "product_constraints": [],
+      "retry_margin": 1.1,
+      "projected_worst_case_usd": 0.013526,
+      "ceiling_usd": 1.0,
+      "within_ceiling": true
+    },
+    {
+      "lane": "mem0-oss",
+      "worst_case_usd": 0.11353,
+      "token_totals": {
+        "chat_input_tokens": 186010,
+        "chat_output_tokens": 18000,
+        "embed_tokens": 239
+      },
+      "line_count": 66,
+      "product_constraints": [],
+      "retry_margin": 1.1,
+      "projected_worst_case_usd": 0.11353,
+      "ceiling_usd": 1.0,
+      "within_ceiling": true
+    },
+    {
+      "lane": "zep",
+      "worst_case_usd": 0.0,
+      "token_totals": {
+        "chat_input_tokens": 1122,
+        "chat_output_tokens": 0,
+        "embed_tokens": 0
+      },
+      "line_count": 47,
+      "product_constraints": [
+        "Zep Cloud native extraction; OpenAI pins not applied; $0 paid ceiling."
+      ],
+      "retry_margin": 1.1,
+      "projected_worst_case_usd": 0.0,
+      "ceiling_usd": 0.0,
+      "within_ceiling": true
+    },
+    {
+      "lane": "_total_openai_worst_case",
+      "projected_worst_case_usd": 0.144592,
+      "ceiling_usd": 3.0,
+      "within_ceiling": true
+    }
+  ],
+  "methodology": "tiktoken cl100k_base on serialized request bodies incl. mem0 ADDITIVE_EXTRACTION_PROMPT; conservative max output caps; 1.10 retry margin; Zep excluded from OpenAI totals."
+}
diff --git a/docs/benchmarks/artifacts/opmem-external-lanes/dry-run-results.json b/docs/benchmarks/artifacts/opmem-external-lanes/dry-run-results.json
new file mode 100644
index 00000000..be0ab0bf
--- /dev/null
+++ b/docs/benchmarks/artifacts/opmem-external-lanes/dry-run-results.json
@@ -0,0 +1,453 @@
+{
+  "benchmark": "opmem-v0",
+  "mode": "dry_run",
+  "run_id": "opmem-dryrun-0001",
+  "task_count": 13,
+  "step_count": 47,
+  "lanes": {
+    "langmem": {
+      "infrastructure_errors": 0,
+      "results": [
+        {
+          "task": "cor01_basic_revision",
+          "outcome": {
+            "passed": false,
+            "failures": [
+              "step1(recall): expected at least 1 results, got 0",
+              "step1(recall): top result does not contain 'python'"
+            ]
+          }
+        },
+        {
+          "task": "cor02_correction_stickiness",
+          "outcome": {
+            "passed": false,
+            "failures": [
+              "step3(recall): expected at least 1 results, got 0",
+              "step3(recall): top result does not contain 'ruby'"
+            ]
+          }
+        },
+        {
+          "task": "cor03_revised_retrievable",
+          "outcome": {
+            "passed": true,
+            "failures": []
+          }
+        },
+        {
+          "task": "dup01_idempotent_remember",
+          "outcome": {
+            "passed": true,
+            "failures": []
+          }
+        },
+        {
+          "task": "iso01_subject_isolation",
+          "outcome": {
+            "passed": true,
+            "failures": []
+          }
+        },
+        {
+          "task": "iso02_tenant_isolation",
+          "outcome": {
+            "passed": true,
+            "failures": []
+          }
+        },
+        {
+          "task": "iso03_forget_isolated",
+          "outcome": {
+            "passed": true,
+            "failures": []
+          }
+        },
+        {
+          "task": "sup01_basic_forget",
+          "outcome": {
+            "passed": true,
+            "failures": []
+          }
+        },
+        {
+          "task": "sup02_targeted_forget",
+          "outcome": {
+            "passed": true,
+            "failures": []
+          }
+        },
+        {
+          "task": "sup03_durable_forget",
+          "outcome": {
+            "passed": false,
+            "failures": [
+              "step3(recall): expected at most 0 results, got 1"
+            ]
+          }
+        },
+        {
+          "task": "upd01_stale_fact",
+          "outcome": {
+            "passed": false,
+            "failures": [
+              "step2(recall): top result does not contain 'june'"
+            ]
+          }
+        },
+        {
+          "task": "upd02_preference_change",
+          "outcome": {
+            "passed": false,
+            "failures": [
+              "step2(recall): expected at least 1 results, got 0",
+              "step2(recall): top result does not contain 'sms'"
+            ]
+          }
+        },
+        {
+          "task": "upd03_state_supersession",
+          "outcome": {
+            "passed": false,
+            "failures": [
+              "step2(recall): top result does not contain '2222'"
+            ]
+          }
+        }
+      ]
+    },
+    "letta": {
+      "infrastructure_errors": 0,
+      "results": [
+        {
+          "task": "cor01_basic_revision",
+          "outcome": {
+            "passed": false,
+            "failures": [
+              "step1(recall): expected at least 1 results, got 0",
+              "step1(recall): top result does not contain 'python'"
+            ]
+          }
+        },
+        {
+          "task": "cor02_correction_stickiness",
+          "outcome": {
+            "passed": false,
+            "failures": [
+              "step3(recall): expected at least 1 results, got 0",
+              "step3(recall): top result does not contain 'ruby'"
+            ]
+          }
+        },
+        {
+          "task": "cor03_revised_retrievable",
+          "outcome": {
+            "passed": true,
+            "failures": []
+          }
+        },
+        {
+          "task": "dup01_idempotent_remember",
+          "outcome": {
+            "passed": true,
+            "failures": []
+          }
+        },
+        {
+          "task": "iso01_subject_isolation",
+          "outcome": {
+            "passed": true,
+            "failures": []
+          }
+        },
+        {
+          "task": "iso02_tenant_isolation",
+          "outcome": {
+            "passed": true,
+            "failures": []
+          }
+        },
+        {
+          "task": "iso03_forget_isolated",
+          "outcome": {
+            "passed": true,
+            "failures": []
+          }
+        },
+        {
+          "task": "sup01_basic_forget",
+          "outcome": {
+            "passed": true,
+            "failures": []
+          }
+        },
+        {
+          "task": "sup02_targeted_forget",
+          "outcome": {
+            "passed": true,
+            "failures": []
+          }
+        },
+        {
+          "task": "sup03_durable_forget",
+          "outcome": {
+            "passed": false,
+            "failures": [
+              "step3(recall): expected at most 0 results, got 1"
+            ]
+          }
+        },
+        {
+          "task": "upd01_stale_fact",
+          "outcome": {
+            "passed": false,
+            "failures": [
+              "step2(recall): top result does not contain 'june'"
+            ]
+          }
+        },
+        {
+          "task": "upd02_preference_change",
+          "outcome": {
+            "passed": false,
+            "failures": [
+              "step2(recall): expected at least 1 results, got 0",
+              "step2(recall): top result does not contain 'sms'"
+            ]
+          }
+        },
+        {
+          "task": "upd03_state_supersession",
+          "outcome": {
+            "passed": false,
+            "failures": [
+              "step2(recall): top result does not contain '2222'"
+            ]
+          }
+        }
+      ]
+    },
+    "mem0-oss": {
+      "infrastructure_errors": 0,
+      "results": [
+        {
+          "task": "cor01_basic_revision",
+          "outcome": {
+            "passed": false,
+            "failures": [
+              "step1(recall): expected at least 1 results, got 0",
+              "step1(recall): top result does not contain 'python'"
+            ]
+          }
+        },
+        {
+          "task": "cor02_correction_stickiness",
+          "outcome": {
+            "passed": false,
+            "failures": [
+              "step3(recall): expected at least 1 results, got 0",
+              "step3(recall): top result does not contain 'ruby'"
+            ]
+          }
+        },
+        {
+          "task": "cor03_revised_retrievable",
+          "outcome": {
+            "passed": true,
+            "failures": []
+          }
+        },
+        {
+          "task": "dup01_idempotent_remember",
+          "outcome": {
+            "passed": true,
+            "failures": []
+          }
+        },
+        {
+          "task": "iso01_subject_isolation",
+          "outcome": {
+            "passed": true,
+            "failures": []
+          }
+        },
+        {
+          "task": "iso02_tenant_isolation",
+          "outcome": {
+            "passed": true,
+            "failures": []
+          }
+        },
+        {
+          "task": "iso03_forget_isolated",
+          "outcome": {
+            "passed": true,
+            "failures": []
+          }
+        },
+        {
+          "task": "sup01_basic_forget",
+          "outcome": {
+            "passed": true,
+            "failures": []
+          }
+        },
+        {
+          "task": "sup02_targeted_forget",
+          "outcome": {
+            "passed": true,
+            "failures": []
+          }
+        },
+        {
+          "task": "sup03_durable_forget",
+          "outcome": {
+            "passed": false,
+            "failures": [
+              "step3(recall): expected at most 0 results, got 1"
+            ]
+          }
+        },
+        {
+          "task": "upd01_stale_fact",
+          "outcome": {
+            "passed": false,
+            "failures": [
+              "step2(recall): top result does not contain 'june'"
+            ]
+          }
+        },
+        {
+          "task": "upd02_preference_change",
+          "outcome": {
+            "passed": false,
+            "failures": [
+              "step2(recall): expected at least 1 results, got 0",
+              "step2(recall): top result does not contain 'sms'"
+            ]
+          }
+        },
+        {
+          "task": "upd03_state_supersession",
+          "outcome": {
+            "passed": false,
+            "failures": [
+              "step2(recall): top result does not contain '2222'"
+            ]
+          }
+        }
+      ]
+    },
+    "zep": {
+      "infrastructure_errors": 0,
+      "results": [
+        {
+          "task": "cor01_basic_revision",
+          "outcome": {
+            "passed": false,
+            "failures": [
+              "step1(recall): expected at least 1 results, got 0",
+              "step1(recall): top result does not contain 'python'"
+            ]
+          }
+        },
+        {
+          "task": "cor02_correction_stickiness",
+          "outcome": {
+            "passed": false,
+            "failures": [
+              "step3(recall): expected at least 1 results, got 0",
+              "step3(recall): top result does not contain 'ruby'"
+            ]
+          }
+        },
+        {
+          "task": "cor03_revised_retrievable",
+          "outcome": {
+            "passed": true,
+            "failures": []
+          }
+        },
+        {
+          "task": "dup01_idempotent_remember",
+          "outcome": {
+            "passed": true,
+            "failures": []
+          }
+        },
+        {
+          "task": "iso01_subject_isolation",
+          "outcome": {
+            "passed": true,
+            "failures": []
+          }
+        },
+        {
+          "task": "iso02_tenant_isolation",
+          "outcome": {
+            "passed": true,
+            "failures": []
+          }
+        },
+        {
+          "task": "iso03_forget_isolated",
+          "outcome": {
+            "passed": true,
+            "failures": []
+          }
+        },
+        {
+          "task": "sup01_basic_forget",
+          "outcome": {
+            "passed": true,
+            "failures": []
+          }
+        },
+        {
+          "task": "sup02_targeted_forget",
+          "outcome": {
+            "passed": true,
+            "failures": []
+          }
+        },
+        {
+          "task": "sup03_durable_forget",
+          "outcome": {
+            "passed": false,
+            "failures": [
+              "step3(recall): expected at most 0 results, got 1"
+            ]
+          }
+        },
+        {
+          "task": "upd01_stale_fact",
+          "outcome": {
+            "passed": false,
+            "failures": [
+              "step2(recall): top result does not contain 'june'"
+            ]
+          }
+        },
+        {
+          "task": "upd02_preference_change",
+          "outcome": {
+            "passed": false,
+            "failures": [
+              "step2(recall): expected at least 1 results, got 0",
+              "step2(recall): top result does not contain 'sms'"
+            ]
+          }
+        },
+        {
+          "task": "upd03_state_supersession",
+          "outcome": {
+            "passed": false,
+            "failures": [
+              "step2(recall): top result does not contain '2222'"
+            ]
+          }
+        }
+      ]
+    }
+  }
+}
diff --git a/docs/benchmarks/artifacts/opmem-external-lanes/live-full-opmem-phase2-full-mem0.json b/docs/benchmarks/artifacts/opmem-external-lanes/live-full-opmem-phase2-full-mem0.json
new file mode 100644
index 00000000..f8d538bf
--- /dev/null
+++ b/docs/benchmarks/artifacts/opmem-external-lanes/live-full-opmem-phase2-full-mem0.json
@@ -0,0 +1,88 @@
+[
+  {
+    "lane": "mem0-oss",
+    "status": "completed",
+    "mode": "full",
+    "run_id": "opmem-phase2-full-mem0",
+    "elapsed_seconds": 36.01,
+    "infrastructure_errors": 0,
+    "tasks_run": 13,
+    "tasks_passed": 9,
+    "task_results": [
+      {
+        "task": "cor01_basic_revision",
+        "passed": true,
+        "failures": []
+      },
+      {
+        "task": "cor02_correction_stickiness",
+        "passed": true,
+        "failures": []
+      },
+      {
+        "task": "cor03_revised_retrievable",
+        "passed": true,
+        "failures": []
+      },
+      {
+        "task": "dup01_idempotent_remember",
+        "passed": true,
+        "failures": []
+      },
+      {
+        "task": "iso01_subject_isolation",
+        "passed": true,
+        "failures": []
+      },
+      {
+        "task": "iso02_tenant_isolation",
+        "passed": true,
+        "failures": []
+      },
+      {
+        "task": "iso03_forget_isolated",
+        "passed": true,
+        "failures": []
+      },
+      {
+        "task": "sup01_basic_forget",
+        "passed": true,
+        "failures": []
+      },
+      {
+        "task": "sup02_targeted_forget",
+        "passed": true,
+        "failures": []
+      },
+      {
+        "task": "sup03_durable_forget",
+        "passed": false,
+        "failures": [
+          "step3(recall): expected at most 0 results, got 1"
+        ]
+      },
+      {
+        "task": "upd01_stale_fact",
+        "passed": false,
+        "failures": [
+          "step2(recall): top result does not contain 'june'"
+        ]
+      },
+      {
+        "task": "upd02_preference_change",
+        "passed": false,
+        "failures": [
+          "step2(recall): top result does not contain 'sms'"
+        ]
+      },
+      {
+        "task": "upd03_state_supersession",
+        "passed": false,
+        "failures": [
+          "step2(recall): top result does not contain '2222'"
+        ]
+      }
+    ],
+    "product_constraints": []
+  }
+]
diff --git a/docs/benchmarks/artifacts/opmem-external-lanes/live-full-opmem-phase2-full1.json b/docs/benchmarks/artifacts/opmem-external-lanes/live-full-opmem-phase2-full1.json
new file mode 100644
index 00000000..56345957
--- /dev/null
+++ b/docs/benchmarks/artifacts/opmem-external-lanes/live-full-opmem-phase2-full1.json
@@ -0,0 +1,176 @@
+[
+  {
+    "lane": "mem0-oss",
+    "status": "invalid_incomplete",
+    "mode": "full",
+    "run_id": "opmem-phase2-full1",
+    "elapsed_seconds": 28.41,
+    "infrastructure_errors": 3,
+    "tasks_run": 13,
+    "tasks_passed": 6,
+    "task_results": [
+      {
+        "task": "cor01_basic_revision",
+        "error": "AttributeError: 'dict' object has no attribute 'replace'",
+        "invalid_incomplete": true
+      },
+      {
+        "task": "cor02_correction_stickiness",
+        "error": "AttributeError: 'dict' object has no attribute 'replace'",
+        "invalid_incomplete": true
+      },
+      {
+        "task": "cor03_revised_retrievable",
+        "error": "AttributeError: 'dict' object has no attribute 'replace'",
+        "invalid_incomplete": true
+      },
+      {
+        "task": "dup01_idempotent_remember",
+        "passed": true,
+        "failures": []
+      },
+      {
+        "task": "iso01_subject_isolation",
+        "passed": true,
+        "failures": []
+      },
+      {
+        "task": "iso02_tenant_isolation",
+        "passed": true,
+        "failures": []
+      },
+      {
+        "task": "iso03_forget_isolated",
+        "passed": true,
+        "failures": []
+      },
+      {
+        "task": "sup01_basic_forget",
+        "passed": true,
+        "failures": []
+      },
+      {
+        "task": "sup02_targeted_forget",
+        "passed": true,
+        "failures": []
+      },
+      {
+        "task": "sup03_durable_forget",
+        "passed": false,
+        "failures": [
+          "step3(recall): expected at most 0 results, got 1"
+        ]
+      },
+      {
+        "task": "upd01_stale_fact",
+        "passed": false,
+        "failures": [
+          "step2(recall): top result does not contain 'june'"
+        ]
+      },
+      {
+        "task": "upd02_preference_change",
+        "passed": false,
+        "failures": [
+          "step2(recall): top result does not contain 'sms'"
+        ]
+      },
+      {
+        "task": "upd03_state_supersession",
+        "passed": false,
+        "failures": [
+          "step2(recall): top result does not contain '2222'"
+        ]
+      }
+    ],
+    "product_constraints": []
+  },
+  {
+    "lane": "langmem",
+    "status": "completed",
+    "mode": "full",
+    "run_id": "opmem-phase2-full1",
+    "elapsed_seconds": 42.07,
+    "infrastructure_errors": 0,
+    "tasks_run": 13,
+    "tasks_passed": 8,
+    "task_results": [
+      {
+        "task": "cor01_basic_revision",
+        "passed": true,
+        "failures": []
+      },
+      {
+        "task": "cor02_correction_stickiness",
+        "passed": false,
+        "failures": [
+          "step3(recall): top result does not contain 'ruby'"
+        ]
+      },
+      {
+        "task": "cor03_revised_retrievable",
+        "passed": true,
+        "failures": []
+      },
+      {
+        "task": "dup01_idempotent_remember",
+        "passed": true,
+        "failures": []
+      },
+      {
+        "task": "iso01_subject_isolation",
+        "passed": true,
+        "failures": []
+      },
+      {
+        "task": "iso02_tenant_isolation",
+        "passed": true,
+        "failures": []
+      },
+      {
+        "task": "iso03_forget_isolated",
+        "passed": false,
+        "failures": [
+          "step3(recall): expected at most 0 results, got 1"
+        ]
+      },
+      {
+        "task": "sup01_basic_forget",
+        "passed": false,
+        "failures": [
+          "step3(recall): expected at most 0 results, got 1"
+        ]
+      },
+      {
+        "task": "sup02_targeted_forget",
+        "passed": false,
+        "failures": [
+          "step3(recall): a result contains excluded 'door code'"
+        ]
+      },
+      {
+        "task": "sup03_durable_forget",
+        "passed": false,
+        "failures": [
+          "step3(recall): expected at most 0 results, got 1"
+        ]
+      },
+      {
+        "task": "upd01_stale_fact",
+        "passed": true,
+        "failures": []
+      },
+      {
+        "task": "upd02_preference_change",
+        "passed": true,
+        "failures": []
+      },
+      {
+        "task": "upd03_state_supersession",
+        "passed": true,
+        "failures": []
+      }
+    ],
+    "product_constraints": []
+  }
+]
diff --git a/docs/benchmarks/artifacts/opmem-external-lanes/live-full-opmem-phase2-letta-full.json b/docs/benchmarks/artifacts/opmem-external-lanes/live-full-opmem-phase2-letta-full.json
new file mode 100644
index 00000000..dfb5f711
--- /dev/null
+++ b/docs/benchmarks/artifacts/opmem-external-lanes/live-full-opmem-phase2-letta-full.json
@@ -0,0 +1,99 @@
+[
+  {
+    "lane": "letta",
+    "status": "completed",
+    "mode": "full",
+    "run_id": "opmem-phase2-letta-full",
+    "elapsed_seconds": 6.56,
+    "infrastructure_errors": 0,
+    "tasks_run": 13,
+    "tasks_passed": 5,
+    "task_results": [
+      {
+        "task": "cor01_basic_revision",
+        "passed": false,
+        "failures": [
+          "step1(recall): expected at least 1 results, got 0",
+          "step1(recall): top result does not contain 'python'"
+        ]
+      },
+      {
+        "task": "cor02_correction_stickiness",
+        "passed": false,
+        "failures": [
+          "step3(recall): expected at least 1 results, got 0",
+          "step3(recall): top result does not contain 'ruby'"
+        ]
+      },
+      {
+        "task": "cor03_revised_retrievable",
+        "passed": true,
+        "failures": []
+      },
+      {
+        "task": "dup01_idempotent_remember",
+        "passed": true,
+        "failures": []
+      },
+      {
+        "task": "iso01_subject_isolation",
+        "passed": false,
+        "failures": [
+          "step1(recall): expected at most 0 results, got 1"
+        ]
+      },
+      {
+        "task": "iso02_tenant_isolation",
+        "passed": false,
+        "failures": [
+          "step1(recall): expected at most 0 results, got 1"
+        ]
+      },
+      {
+        "task": "iso03_forget_isolated",
+        "passed": false,
+        "failures": [
+          "step4(recall): expected at least 1 results, got 0"
+        ]
+      },
+      {
+        "task": "sup01_basic_forget",
+        "passed": true,
+        "failures": []
+      },
+      {
+        "task": "sup02_targeted_forget",
+        "passed": true,
+        "failures": []
+      },
+      {
+        "task": "sup03_durable_forget",
+        "passed": false,
+        "failures": [
+          "step3(recall): expected at most 0 results, got 1"
+        ]
+      },
+      {
+        "task": "upd01_stale_fact",
+        "passed": false,
+        "failures": [
+          "step2(recall): top result does not contain 'june'"
+        ]
+      },
+      {
+        "task": "upd02_preference_change",
+        "passed": false,
+        "failures": [
+          "step2(recall): expected at least 1 results, got 0",
+          "step2(recall): top result does not contain 'sms'"
+        ]
+      },
+      {
+        "task": "upd03_state_supersession",
+        "passed": true,
+        "failures": []
+      }
+    ],
+    "product_constraints": []
+  }
+]
diff --git a/docs/benchmarks/artifacts/opmem-external-lanes/live-full-opmem-phase2-zep-full.json b/docs/benchmarks/artifacts/opmem-external-lanes/live-full-opmem-phase2-zep-full.json
new file mode 100644
index 00000000..ccf6e689
--- /dev/null
+++ b/docs/benchmarks/artifacts/opmem-external-lanes/live-full-opmem-phase2-zep-full.json
@@ -0,0 +1,106 @@
+[
+  {
+    "lane": "zep",
+    "status": "completed",
+    "mode": "full",
+    "run_id": "opmem-phase2-zep-full",
+    "elapsed_seconds": 51.01,
+    "infrastructure_errors": 0,
+    "tasks_run": 13,
+    "tasks_passed": 3,
+    "task_results": [
+      {
+        "task": "cor01_basic_revision",
+        "passed": false,
+        "failures": [
+          "step1(recall): expected at least 1 results, got 0",
+          "step1(recall): top result does not contain 'python'"
+        ]
+      },
+      {
+        "task": "cor02_correction_stickiness",
+        "passed": true,
+        "failures": []
+      },
+      {
+        "task": "cor03_revised_retrievable",
+        "passed": true,
+        "failures": []
+      },
+      {
+        "task": "dup01_idempotent_remember",
+        "passed": true,
+        "failures": []
+      },
+      {
+        "task": "iso01_subject_isolation",
+        "passed": false,
+        "failures": [
+          "step1(recall): expected at most 0 results, got 1"
+        ]
+      },
+      {
+        "task": "iso02_tenant_isolation",
+        "passed": false,
+        "failures": [
+          "step1(recall): expected at most 0 results, got 1"
+        ]
+      },
+      {
+        "task": "iso03_forget_isolated",
+        "passed": false,
+        "failures": [
+          "step3(recall): expected at most 0 results, got 1"
+        ]
+      },
+      {
+        "task": "sup01_basic_forget",
+        "passed": false,
+        "failures": [
+          "step3(recall): expected at most 0 results, got 1"
+        ]
+      },
+      {
+        "task": "sup02_targeted_forget",
+        "passed": false,
+        "failures": [
+          "step3(recall): a result contains excluded 'door code'"
+        ]
+      },
+      {
+        "task": "sup03_durable_forget",
+        "passed": false,
+        "failures": [
+          "step3(recall): expected at most 0 results, got 2"
+        ]
+      },
+      {
+        "task": "upd01_stale_fact",
+        "passed": false,
+        "failures": [
+          "step2(recall): top result does not contain 'june'"
+        ]
+      },
+      {
+        "task": "upd02_preference_change",
+        "passed": false,
+        "failures": [
+          "step2(recall): expected at least 1 results, got 0",
+          "step2(recall): top result does not contain 'sms'"
+        ]
+      },
+      {
+        "task": "upd03_state_supersession",
+        "passed": false,
+        "failures": [
+          "step2(recall): top result does not contain '2222'"
+        ]
+      }
+    ],
+    "product_constraints": [
+      "lane zep cannot pin chat model gpt-4.1-mini-2025-04-14",
+      "lane zep cannot pin embed model text-embedding-3-small",
+      "Zep Cloud memory path does not accept OpenAI chat/embed model pins; stop before substituting \u2014 record scores only under Zep-native models."
+    ]
+  }
+]
diff --git a/docs/benchmarks/artifacts/opmem-external-lanes/live-smoke-opmem-phase2-letta-smoke.json b/docs/benchmarks/artifacts/opmem-external-lanes/live-smoke-opmem-phase2-letta-smoke.json
new file mode 100644
index 00000000..20d10b9e
--- /dev/null
+++ b/docs/benchmarks/artifacts/opmem-external-lanes/live-smoke-opmem-phase2-letta-smoke.json
@@ -0,0 +1,20 @@
+[
+  {
+    "lane": "letta",
+    "status": "completed",
+    "mode": "smoke",
+    "run_id": "opmem-phase2-letta-smoke",
+    "elapsed_seconds": 0.46,
+    "infrastructure_errors": 0,
+    "tasks_run": 1,
+    "tasks_passed": 1,
+    "task_results": [
+      {
+        "task": "dup01_idempotent_remember",
+        "passed": true,
+        "failures": []
+      }
+    ],
+    "product_constraints": []
+  }
+]
diff --git a/docs/benchmarks/artifacts/opmem-external-lanes/live-smoke-opmem-phase2-smoke1.json b/docs/benchmarks/artifacts/opmem-external-lanes/live-smoke-opmem-phase2-smoke1.json
new file mode 100644
index 00000000..1a3af060
--- /dev/null
+++ b/docs/benchmarks/artifacts/opmem-external-lanes/live-smoke-opmem-phase2-smoke1.json
@@ -0,0 +1,52 @@
+[
+  {
+    "lane": "mem0-oss",
+    "status": "invalid_incomplete",
+    "mode": "smoke",
+    "run_id": "opmem-phase2-smoke1",
+    "elapsed_seconds": 6.36,
+    "infrastructure_errors": 1,
+    "tasks_run": 1,
+    "tasks_passed": 0,
+    "task_results": [
+      {
+        "task": "dup01_idempotent_remember",
+        "error": "ValueError: Top-level entity parameters frozenset({'user_id'}) are not supported in search(). Use filters={'user_id': '...'} instead.",
+        "invalid_incomplete": true
+      }
+    ],
+    "product_constraints": []
+  },
+  {
+    "lane": "langmem",
+    "status": "invalid_incomplete",
+    "mode": "smoke",
+    "run_id": "opmem-phase2-smoke1",
+    "elapsed_seconds": 3.15,
+    "infrastructure_errors": 1,
+    "tasks_run": 1,
+    "tasks_passed": 0,
+    "task_results": [
+      {
+        "task": "dup01_idempotent_remember",
+        "error": "AttributeError: 'dict' object has no attribute 'lower'",
+        "invalid_incomplete": true
+      }
+    ],
+    "product_constraints": []
+  },
+  {
+    "lane": "zep",
+    "status": "blocked",
+    "reason": "ZEP_API_KEY not present in runtime (vault not injected on this VM)",
+    "infrastructure_errors": 0,
+    "task_results": []
+  },
+  {
+    "lane": "letta",
+    "status": "blocked",
+    "reason": "LETTA_APP_SERVER_TOKEN not present in runtime",
+    "infrastructure_errors": 0,
+    "task_results": []
+  }
+]
diff --git a/docs/benchmarks/artifacts/opmem-external-lanes/live-smoke-opmem-phase2-smoke2.json b/docs/benchmarks/artifacts/opmem-external-lanes/live-smoke-opmem-phase2-smoke2.json
new file mode 100644
index 00000000..9b77e193
--- /dev/null
+++ b/docs/benchmarks/artifacts/opmem-external-lanes/live-smoke-opmem-phase2-smoke2.json
@@ -0,0 +1,38 @@
+[
+  {
+    "lane": "mem0-oss",
+    "status": "completed",
+    "mode": "smoke",
+    "run_id": "opmem-phase2-smoke2",
+    "elapsed_seconds": 2.9,
+    "infrastructure_errors": 0,
+    "tasks_run": 1,
+    "tasks_passed": 1,
+    "task_results": [
+      {
+        "task": "dup01_idempotent_remember",
+        "passed": true,
+        "failures": []
+      }
+    ],
+    "product_constraints": []
+  },
+  {
+    "lane": "langmem",
+    "status": "completed",
+    "mode": "smoke",
+    "run_id": "opmem-phase2-smoke2",
+    "elapsed_seconds": 3.68,
+    "infrastructure_errors": 0,
+    "tasks_run": 1,
+    "tasks_passed": 1,
+    "task_results": [
+      {
+        "task": "dup01_idempotent_remember",
+        "passed": true,
+        "failures": []
+      }
+    ],
+    "product_constraints": []
+  }
+]
diff --git a/docs/benchmarks/artifacts/opmem-external-lanes/live-smoke-opmem-phase2-zep-smoke-20260923.json b/docs/benchmarks/artifacts/opmem-external-lanes/live-smoke-opmem-phase2-zep-smoke-20260923.json
new file mode 100644
index 00000000..61b0b2ba
--- /dev/null
+++ b/docs/benchmarks/artifacts/opmem-external-lanes/live-smoke-opmem-phase2-zep-smoke-20260923.json
@@ -0,0 +1,9 @@
+[
+  {
+    "lane": "zep",
+    "status": "blocked",
+    "reason": "ZEP_API_KEY not present in runtime (vault not injected on this VM)",
+    "infrastructure_errors": 0,
+    "task_results": []
+  }
+]
diff --git a/docs/benchmarks/artifacts/opmem-external-lanes/live-smoke-opmem-phase2-zep-smoke-20260923b.json b/docs/benchmarks/artifacts/opmem-external-lanes/live-smoke-opmem-phase2-zep-smoke-20260923b.json
new file mode 100644
index 00000000..08a39ade
--- /dev/null
+++ b/docs/benchmarks/artifacts/opmem-external-lanes/live-smoke-opmem-phase2-zep-smoke-20260923b.json
@@ -0,0 +1,24 @@
+[
+  {
+    "lane": "zep",
+    "status": "invalid_incomplete",
+    "mode": "smoke",
+    "run_id": "opmem-phase2-zep-smoke-20260923b",
+    "elapsed_seconds": 0.32,
+    "infrastructure_errors": 1,
+    "tasks_run": 1,
+    "tasks_passed": 0,
+    "task_results": [
+      {
+        "task": "dup01_idempotent_remember",
+        "error": "AttributeError: 'Zep' object has no attribute 'memory'",
+        "invalid_incomplete": true
+      }
+    ],
+    "product_constraints": [
+      "lane zep cannot pin chat model gpt-4.1-mini-2025-04-14",
+      "lane zep cannot pin embed model text-embedding-3-small",
+      "Zep Cloud memory path does not accept OpenAI chat/embed model pins; stop before substituting \u2014 record scores only under Zep-native models."
+    ]
+  }
+]
diff --git a/docs/benchmarks/artifacts/opmem-external-lanes/live-smoke-opmem-phase2-zep-smoke-20260923c.json b/docs/benchmarks/artifacts/opmem-external-lanes/live-smoke-opmem-phase2-zep-smoke-20260923c.json
new file mode 100644
index 00000000..51fea514
--- /dev/null
+++ b/docs/benchmarks/artifacts/opmem-external-lanes/live-smoke-opmem-phase2-zep-smoke-20260923c.json
@@ -0,0 +1,26 @@
+[
+  {
+    "lane": "zep",
+    "status": "completed",
+    "mode": "smoke",
+    "run_id": "opmem-phase2-zep-smoke-20260923c",
+    "elapsed_seconds": 3.95,
+    "infrastructure_errors": 0,
+    "tasks_run": 1,
+    "tasks_passed": 0,
+    "task_results": [
+      {
+        "task": "dup01_idempotent_remember",
+        "passed": false,
+        "failures": [
+          "step2(recall): expected at least 1 results, got 0"
+        ]
+      }
+    ],
+    "product_constraints": [
+      "lane zep cannot pin chat model gpt-4.1-mini-2025-04-14",
+      "lane zep cannot pin embed model text-embedding-3-small",
+      "Zep Cloud memory path does not accept OpenAI chat/embed model pins; stop before substituting \u2014 record scores only under Zep-native models."
+    ]
+  }
+]
diff --git a/docs/benchmarks/artifacts/opmem-external-lanes/live-smoke-opmem-phase2-zep-smoke-final.json b/docs/benchmarks/artifacts/opmem-external-lanes/live-smoke-opmem-phase2-zep-smoke-final.json
new file mode 100644
index 00000000..eb9c1494
--- /dev/null
+++ b/docs/benchmarks/artifacts/opmem-external-lanes/live-smoke-opmem-phase2-zep-smoke-final.json
@@ -0,0 +1,26 @@
+[
+  {
+    "lane": "zep",
+    "status": "completed",
+    "mode": "smoke",
+    "run_id": "opmem-phase2-zep-smoke-final",
+    "elapsed_seconds": 4.01,
+    "infrastructure_errors": 0,
+    "tasks_run": 1,
+    "tasks_passed": 0,
+    "task_results": [
+      {
+        "task": "dup01_idempotent_remember",
+        "passed": false,
+        "failures": [
+          "step2(recall): expected at most 1 results, got 2"
+        ]
+      }
+    ],
+    "product_constraints": [
+      "lane zep cannot pin chat model gpt-4.1-mini-2025-04-14",
+      "lane zep cannot pin embed model text-embedding-3-small",
+      "Zep Cloud memory path does not accept OpenAI chat/embed model pins; stop before substituting \u2014 record scores only under Zep-native models."
+    ]
+  }
+]
diff --git a/docs/benchmarks/artifacts/opmem-external-lanes/live-smoke-opmem-phase2-zep-smoke.json b/docs/benchmarks/artifacts/opmem-external-lanes/live-smoke-opmem-phase2-zep-smoke.json
new file mode 100644
index 00000000..61b0b2ba
--- /dev/null
+++ b/docs/benchmarks/artifacts/opmem-external-lanes/live-smoke-opmem-phase2-zep-smoke.json
@@ -0,0 +1,9 @@
+[
+  {
+    "lane": "zep",
+    "status": "blocked",
+    "reason": "ZEP_API_KEY not present in runtime (vault not injected on this VM)",
+    "infrastructure_errors": 0,
+    "task_results": []
+  }
+]
diff --git a/docs/benchmarks/artifacts/opmem-external-lanes/live-zep-blocked-opmem-phase2-zep-smoke.json b/docs/benchmarks/artifacts/opmem-external-lanes/live-zep-blocked-opmem-phase2-zep-smoke.json
new file mode 100644
index 00000000..64a87cbd
--- /dev/null
+++ b/docs/benchmarks/artifacts/opmem-external-lanes/live-zep-blocked-opmem-phase2-zep-smoke.json
@@ -0,0 +1,11 @@
+{
+  "lane": "zep",
+  "status": "blocked",
+  "reason": "ZEP_API_KEY not present in runtime (not in CLOUD_AGENT_INJECTED_SECRET_NAMES on 2026-09-21 cloud VM)",
+  "run_id": "opmem-phase2-zep-smoke",
+  "preflight": {
+    "free_tier_check": "not_attempted",
+    "paid_ceiling_usd": 0.0
+  },
+  "reproduce": "export ZEP_API_KEY=<free-tier key>; python3 evals/run_opmem_external_live.py --smoke --lanes zep"
+}
diff --git a/docs/benchmarks/artifacts/opmem-external-lanes/raw-langmem-dry-run.json b/docs/benchmarks/artifacts/opmem-external-lanes/raw-langmem-dry-run.json
new file mode 100644
index 00000000..b608b4a0
--- /dev/null
+++ b/docs/benchmarks/artifacts/opmem-external-lanes/raw-langmem-dry-run.json
@@ -0,0 +1,331 @@
+[
+  {
+    "lane": "langmem",
+    "task": "cor01_basic_revision",
+    "op": "remember",
+    "status": 200,
+    "body_preview": "{\"memory_id\": \"m-ec287521fe29\", \"dry_run\": true}"
+  },
+  {
+    "lane": "langmem",
+    "task": "cor01_basic_revision",
+    "op": "recall",
+    "status": 200,
+    "body_preview": "{\"results\": [], \"dry_run\": true}"
+  },
+  {
+    "lane": "langmem",
+    "task": "cor01_basic_revision",
+    "op": "revise",
+    "status": 200,
+    "body_preview": "{\"memory_id\": [\"m-ec287521fe29\"], \"dry_run\": true}"
+  },
+  {
+    "lane": "langmem",
+    "task": "cor01_basic_revision",
+    "op": "recall",
+    "status": 200,
+    "body_preview": "{\"results\": [{\"id\": \"m-ec287521fe29\", \"content\": \"Prefers Ruby for scripting\"}], \"dry_run\": true}"
+  },
+  {
+    "lane": "langmem",
+    "task": "cor02_correction_stickiness",
+    "op": "remember",
+    "status": 200,
+    "body_preview": "{\"memory_id\": \"m-ec77cf3747d2\", \"dry_run\": true}"
+  },
+  {
+    "lane": "langmem",
+    "task": "cor02_correction_stickiness",
+    "op": "revise",
+    "status": 200,
+    "body_preview": "{\"memory_id\": [\"m-ec77cf3747d2\"], \"dry_run\": true}"
+  },
+  {
+    "lane": "langmem",
+    "task": "cor02_correction_stickiness",
+    "op": "remember",
+    "status": 200,
+    "body_preview": "{\"memory_id\": \"m-ec77cf3747d2\", \"dry_run\": true}"
+  },
+  {
+    "lane": "langmem",
+    "task": "cor02_correction_stickiness",
+    "op": "recall",
+    "status": 200,
+    "body_preview": "{\"results\": [], \"dry_run\": true}"
+  },
+  {
+    "lane": "langmem",
+    "task": "cor03_revised_retrievable",
+    "op": "remember",
+    "status": 200,
+    "body_preview": "{\"memory_id\": \"m-97e4481c7620\", \"dry_run\": true}"
+  },
+  {
+    "lane": "langmem",
+    "task": "cor03_revised_retrievable",
+    "op": "revise",
+    "status": 200,
+    "body_preview": "{\"memory_id\": [\"m-97e4481c7620\"], \"dry_run\": true}"
+  },
+  {
+    "lane": "langmem",
+    "task": "cor03_revised_retrievable",
+    "op": "recall",
+    "status": 200,
+    "body_preview": "{\"results\": [{\"id\": \"m-97e4481c7620\", \"content\": \"Prefers Ruby for scripting\"}], \"dry_run\": true}"
+  },
+  {
+    "lane": "langmem",
+    "task": "dup01_idempotent_remember",
+    "op": "remember",
+    "status": 200,
+    "body_preview": "{\"memory_id\": \"m-a2ea413a2ee8\", \"dry_run\": true}"
+  },
+  {
+    "lane": "langmem",
+    "task": "dup01_idempotent_remember",
+    "op": "remember",
+    "status": 200,
+    "body_preview": "{\"memory_id\": \"m-a2ea413a2ee8\", \"dry_run\": true}"
+  },
+  {
+    "lane": "langmem",
+    "task": "dup01_idempotent_remember",
+    "op": "recall",
+    "status": 200,
+    "body_preview": "{\"results\": [{\"id\": \"m-a2ea413a2ee8\", \"content\": \"The wifi network name is FalconNet.\"}], \"dry_run\": true}"
+  },
+  {
+    "lane": "langmem",
+    "task": "iso01_subject_isolation",
+    "op": "remember",
+    "status": 200,
+    "body_preview": "{\"memory_id\": \"m-f1a8b730d634\", \"dry_run\": true}"
+  },
+  {
+    "lane": "langmem",
+    "task": "iso01_subject_isolation",
+    "op": "recall",
+    "status": 200,
+    "body_preview": "{\"results\": [], \"dry_run\": true}"
+  },
+  {
+    "lane": "langmem",
+    "task": "iso01_subject_isolation",
+    "op": "recall",
+    "status": 200,
+    "body_preview": "{\"results\": [{\"id\": \"m-f1a8b730d634\", \"content\": \"My name is Alice Chen.\"}], \"dry_run\": true}"
+  },
+  {
+    "lane": "langmem",
+    "task": "iso02_tenant_isolation",
+    "op": "remember",
+    "status": 200,
+    "body_preview": "{\"memory_id\": \"m-c790561d256a\", \"dry_run\": true}"
+  },
+  {
+    "lane": "langmem",
+    "task": "iso02_tenant_isolation",
+    "op": "recall",
+    "status": 200,
+    "body_preview": "{\"results\": [], \"dry_run\": true}"
+  },
+  {
+    "lane": "langmem",
+    "task": "iso02_tenant_isolation",
+    "op": "recall",
+    "status": 200,
+    "body_preview": "{\"results\": [{\"id\": \"m-c790561d256a\", \"content\": \"The wifi network name is FalconNet.\"}], \"dry_run\": true}"
+  },
+  {
+    "lane": "langmem",
+    "task": "iso03_forget_isolated",
+    "op": "remember",
+    "status": 200,
+    "body_preview": "{\"memory_id\": \"m-b9f3133e0f76\", \"dry_run\": true}"
+  },
+  {
+    "lane": "langmem",
+    "task": "iso03_forget_isolated",
+    "op": "remember",
+    "status": 200,
+    "body_preview": "{\"memory_id\": \"m-b9f3133e0f76\", \"dry_run\": true}"
+  },
+  {
+    "lane": "langmem",
+    "task": "iso03_forget_isolated",
+    "op": "forget",
+    "status": 200,
+    "body_preview": "{\"deleted\": [\"m-b9f3133e0f76\"], \"dry_run\": true}"
+  },
+  {
+    "lane": "langmem",
+    "task": "iso03_forget_isolated",
+    "op": "recall",
+    "status": 200,
+    "body_preview": "{\"results\": [], \"dry_run\": true}"
+  },
+  {
+    "lane": "langmem",
+    "task": "iso03_forget_isolated",
+    "op": "recall",
+    "status": 200,
+    "body_preview": "{\"results\": [{\"id\": \"m-b9f3133e0f76\", \"content\": \"Never share the door code with vendors.\"}], \"dry_run\": true}"
+  },
+  {
+    "lane": "langmem",
+    "task": "sup01_basic_forget",
+    "op": "remember",
+    "status": 200,
+    "body_preview": "{\"memory_id\": \"m-12df728344c1\", \"dry_run\": true}"
+  },
+  {
+    "lane": "langmem",
+    "task": "sup01_basic_forget",
+    "op": "recall",
+    "status": 200,
+    "body_preview": "{\"results\": [{\"id\": \"m-12df728344c1\", \"content\": \"Never share the door code with vendors.\"}], \"dry_run\": true}"
+  },
+  {
+    "lane": "langmem",
+    "task": "sup01_basic_forget",
+    "op": "forget",
+    "status": 200,
+    "body_preview": "{\"deleted\": [\"m-12df728344c1\"], \"dry_run\": true}"
+  },
+  {
+    "lane": "langmem",
+    "task": "sup01_basic_forget",
+    "op": "recall",
+    "status": 200,
+    "body_preview": "{\"results\": [], \"dry_run\": true}"
+  },
+  {
+    "lane": "langmem",
+    "task": "sup02_targeted_forget",
+    "op": "remember",
+    "status": 200,
+    "body_preview": "{\"memory_id\": \"m-694a3ac917b4\", \"dry_run\": true}"
+  },
+  {
+    "lane": "langmem",
+    "task": "sup02_targeted_forget",
+    "op": "remember",
+    "status": 200,
+    "body_preview": "{\"memory_id\": \"m-abc4cd827fe4\", \"dry_run\": true}"
+  },
+  {
+    "lane": "langmem",
+    "task": "sup02_targeted_forget",
+    "op": "forget",
+    "status": 200,
+    "body_preview": "{\"deleted\": [\"m-694a3ac917b4\"], \"dry_run\": true}"
+  },
+  {
+    "lane": "langmem",
+    "task": "sup02_targeted_forget",
+    "op": "recall",
+    "status": 200,
+    "body_preview": "{\"results\": [], \"dry_run\": true}"
+  },
+  {
+    "lane": "langmem",
+    "task": "sup02_targeted_forget",
+    "op": "recall",
+    "status": 200,
+    "body_preview": "{\"results\": [{\"id\": \"m-abc4cd827fe4\", \"content\": \"The wifi network name is FalconNet.\"}], \"dry_run\": true}"
+  },
+  {
+    "lane": "langmem",
+    "task": "sup03_durable_forget",
+    "op": "remember",
+    "status": 200,
+    "body_preview": "{\"memory_id\": \"m-f40e06dae9df\", \"dry_run\": true}"
+  },
+  {
+    "lane": "langmem",
+    "task": "sup03_durable_forget",
+    "op": "forget",
+    "status": 200,
+    "body_preview": "{\"deleted\": [\"m-f40e06dae9df\"], \"dry_run\": true}"
+  },
+  {
+    "lane": "langmem",
+    "task": "sup03_durable_forget",
+    "op": "remember",
+    "status": 200,
+    "body_preview": "{\"memory_id\": \"m-f40e06dae9df\", \"dry_run\": true}"
+  },
+  {
+    "lane": "langmem",
+    "task": "sup03_durable_forget",
+    "op": "recall",
+    "status": 200,
+    "body_preview": "{\"results\": [{\"id\": \"m-f40e06dae9df\", \"content\": \"Never share the door code with vendors.\"}], \"dry_run\": true}"
+  },
+  {
+    "lane": "langmem",
+    "task": "upd01_stale_fact",
+    "op": "remember",
+    "status": 200,
+    "body_preview": "{\"memory_id\": \"m-47fbfe8f74ed\", \"dry_run\": true}"
+  },
+  {
+    "lane": "langmem",
+    "task": "upd01_stale_fact",
+    "op": "remember",
+    "status": 200,
+    "body_preview": "{\"memory_id\": \"m-e5330f7a0060\", \"dry_run\": true}"
+  },
+  {
+    "lane": "langmem",
+    "task": "upd01_stale_fact",
+    "op": "recall",
+    "status": 200,
+    "body_preview": "{\"results\": [{\"id\": \"m-47fbfe8f74ed\", \"content\": \"The launch date is May 12.\"}, {\"id\": \"m-e5330f7a0060\", \"content\": \"The launch date is June 3.\"}], \"dry_run\": true}"
+  },
+  {
+    "lane": "langmem",
+    "task": "upd02_preference_change",
+    "op": "remember",
+    "status": 200,
+    "body_preview": "{\"memory_id\": \"m-e582398e5c23\", \"dry_run\": true}"
+  },
+  {
+    "lane": "langmem",
+    "task": "upd02_preference_change",
+    "op": "remember",
+    "status": 200,
+    "body_preview": "{\"memory_id\": \"m-02ba67213b96\", \"dry_run\": true}"
+  },
+  {
+    "lane": "langmem",
+    "task": "upd02_preference_change",
+    "op": "recall",
+    "status": 200,
+    "body_preview": "{\"results\": [], \"dry_run\": true}"
+  },
+  {
+    "lane": "langmem",
+    "task": "upd03_state_supersession",
+    "op": "remember",
+    "status": 200,
+    "body_preview": "{\"memory_id\": \"m-d0cc03c9eecd\", \"dry_run\": true}"
+  },
+  {
+    "lane": "langmem",
+    "task": "upd03_state_supersession",
+    "op": "remember",
+    "status": 200,
+    "body_preview": "{\"memory_id\": \"m-5230d4b1d371\", \"dry_run\": true}"
+  },
+  {
+    "lane": "langmem",
+    "task": "upd03_state_supersession",
+    "op": "recall",
+    "status": 200,
+    "body_preview": "{\"results\": [{\"id\": \"m-d0cc03c9eecd\", \"content\": \"The door code is 1111.\"}, {\"id\": \"m-5230d4b1d371\", \"content\": \"The door code is 2222.\"}], \"dry_run\": true}"
+  }
+]
diff --git a/docs/benchmarks/artifacts/opmem-external-lanes/raw-langmem-full.json b/docs/benchmarks/artifacts/opmem-external-lanes/raw-langmem-full.json
new file mode 100644
index 00000000..fe51488c
--- /dev/null
+++ b/docs/benchmarks/artifacts/opmem-external-lanes/raw-langmem-full.json
@@ -0,0 +1 @@
+[]
diff --git a/docs/benchmarks/artifacts/opmem-external-lanes/raw-langmem-smoke.json b/docs/benchmarks/artifacts/opmem-external-lanes/raw-langmem-smoke.json
new file mode 100644
index 00000000..fe51488c
--- /dev/null
+++ b/docs/benchmarks/artifacts/opmem-external-lanes/raw-langmem-smoke.json
@@ -0,0 +1 @@
+[]
diff --git a/docs/benchmarks/artifacts/opmem-external-lanes/raw-letta-dry-run.json b/docs/benchmarks/artifacts/opmem-external-lanes/raw-letta-dry-run.json
new file mode 100644
index 00000000..aa2d1273
--- /dev/null
+++ b/docs/benchmarks/artifacts/opmem-external-lanes/raw-letta-dry-run.json
@@ -0,0 +1,331 @@
+[
+  {
+    "lane": "letta",
+    "task": "cor01_basic_revision",
+    "op": "remember",
+    "status": 200,
+    "body_preview": "{\"memory_id\": \"m-ec287521fe29\", \"dry_run\": true}"
+  },
+  {
+    "lane": "letta",
+    "task": "cor01_basic_revision",
+    "op": "recall",
+    "status": 200,
+    "body_preview": "{\"results\": [], \"dry_run\": true}"
+  },
+  {
+    "lane": "letta",
+    "task": "cor01_basic_revision",
+    "op": "revise",
+    "status": 200,
+    "body_preview": "{\"memory_id\": [\"m-ec287521fe29\"], \"dry_run\": true}"
+  },
+  {
+    "lane": "letta",
+    "task": "cor01_basic_revision",
+    "op": "recall",
+    "status": 200,
+    "body_preview": "{\"results\": [{\"id\": \"m-ec287521fe29\", \"content\": \"Prefers Ruby for scripting\"}], \"dry_run\": true}"
+  },
+  {
+    "lane": "letta",
+    "task": "cor02_correction_stickiness",
+    "op": "remember",
+    "status": 200,
+    "body_preview": "{\"memory_id\": \"m-ec77cf3747d2\", \"dry_run\": true}"
+  },
+  {
+    "lane": "letta",
+    "task": "cor02_correction_stickiness",
+    "op": "revise",
+    "status": 200,
+    "body_preview": "{\"memory_id\": [\"m-ec77cf3747d2\"], \"dry_run\": true}"
+  },
+  {
+    "lane": "letta",
+    "task": "cor02_correction_stickiness",
+    "op": "remember",
+    "status": 200,
+    "body_preview": "{\"memory_id\": \"m-ec77cf3747d2\", \"dry_run\": true}"
+  },
+  {
+    "lane": "letta",
+    "task": "cor02_correction_stickiness",
+    "op": "recall",
+    "status": 200,
+    "body_preview": "{\"results\": [], \"dry_run\": true}"
+  },
+  {
+    "lane": "letta",
+    "task": "cor03_revised_retrievable",
+    "op": "remember",
+    "status": 200,
+    "body_preview": "{\"memory_id\": \"m-97e4481c7620\", \"dry_run\": true}"
+  },
+  {
+    "lane": "letta",
+    "task": "cor03_revised_retrievable",
+    "op": "revise",
+    "status": 200,
+    "body_preview": "{\"memory_id\": [\"m-97e4481c7620\"], \"dry_run\": true}"
+  },
+  {
+    "lane": "letta",
+    "task": "cor03_revised_retrievable",
+    "op": "recall",
+    "status": 200,
+    "body_preview": "{\"results\": [{\"id\": \"m-97e4481c7620\", \"content\": \"Prefers Ruby for scripting\"}], \"dry_run\": true}"
+  },
+  {
+    "lane": "letta",
+    "task": "dup01_idempotent_remember",
+    "op": "remember",
+    "status": 200,
+    "body_preview": "{\"memory_id\": \"m-a2ea413a2ee8\", \"dry_run\": true}"
+  },
+  {
+    "lane": "letta",
+    "task": "dup01_idempotent_remember",
+    "op": "remember",
+    "status": 200,
+    "body_preview": "{\"memory_id\": \"m-a2ea413a2ee8\", \"dry_run\": true}"
+  },
+  {
+    "lane": "letta",
+    "task": "dup01_idempotent_remember",
+    "op": "recall",
+    "status": 200,
+    "body_preview": "{\"results\": [{\"id\": \"m-a2ea413a2ee8\", \"content\": \"The wifi network name is FalconNet.\"}], \"dry_run\": true}"
+  },
+  {
+    "lane": "letta",
+    "task": "iso01_subject_isolation",
+    "op": "remember",
+    "status": 200,
+    "body_preview": "{\"memory_id\": \"m-f1a8b730d634\", \"dry_run\": true}"
+  },
+  {
+    "lane": "letta",
+    "task": "iso01_subject_isolation",
+    "op": "recall",
+    "status": 200,
+    "body_preview": "{\"results\": [], \"dry_run\": true}"
+  },
+  {
+    "lane": "letta",
+    "task": "iso01_subject_isolation",
+    "op": "recall",
+    "status": 200,
+    "body_preview": "{\"results\": [{\"id\": \"m-f1a8b730d634\", \"content\": \"My name is Alice Chen.\"}], \"dry_run\": true}"
+  },
+  {
+    "lane": "letta",
+    "task": "iso02_tenant_isolation",
+    "op": "remember",
+    "status": 200,
+    "body_preview": "{\"memory_id\": \"m-c790561d256a\", \"dry_run\": true}"
+  },
+  {
+    "lane": "letta",
+    "task": "iso02_tenant_isolation",
+    "op": "recall",
+    "status": 200,
+    "body_preview": "{\"results\": [], \"dry_run\": true}"
+  },
+  {
+    "lane": "letta",
+    "task": "iso02_tenant_isolation",
+    "op": "recall",
+    "status": 200,
+    "body_preview": "{\"results\": [{\"id\": \"m-c790561d256a\", \"content\": \"The wifi network name is FalconNet.\"}], \"dry_run\": true}"
+  },
+  {
+    "lane": "letta",
+    "task": "iso03_forget_isolated",
+    "op": "remember",
+    "status": 200,
+    "body_preview": "{\"memory_id\": \"m-b9f3133e0f76\", \"dry_run\": true}"
+  },
+  {
+    "lane": "letta",
+    "task": "iso03_forget_isolated",
+    "op": "remember",
+    "status": 200,
+    "body_preview": "{\"memory_id\": \"m-b9f3133e0f76\", \"dry_run\": true}"
+  },
+  {
+    "lane": "letta",
+    "task": "iso03_forget_isolated",
+    "op": "forget",
+    "status": 200,
+    "body_preview": "{\"deleted\": [\"m-b9f3133e0f76\"], \"dry_run\": true}"
+  },
+  {
+    "lane": "letta",
+    "task": "iso03_forget_isolated",
+    "op": "recall",
+    "status": 200,
+    "body_preview": "{\"results\": [], \"dry_run\": true}"
+  },
+  {
+    "lane": "letta",
+    "task": "iso03_forget_isolated",
+    "op": "recall",
+    "status": 200,
+    "body_preview": "{\"results\": [{\"id\": \"m-b9f3133e0f76\", \"content\": \"Never share the door code with vendors.\"}], \"dry_run\": true}"
+  },
+  {
+    "lane": "letta",
+    "task": "sup01_basic_forget",
+    "op": "remember",
+    "status": 200,
+    "body_preview": "{\"memory_id\": \"m-12df728344c1\", \"dry_run\": true}"
+  },
+  {
+    "lane": "letta",
+    "task": "sup01_basic_forget",
+    "op": "recall",
+    "status": 200,
+    "body_preview": "{\"results\": [{\"id\": \"m-12df728344c1\", \"content\": \"Never share the door code with vendors.\"}], \"dry_run\": true}"
+  },
+  {
+    "lane": "letta",
+    "task": "sup01_basic_forget",
+    "op": "forget",
+    "status": 200,
+    "body_preview": "{\"deleted\": [\"m-12df728344c1\"], \"dry_run\": true}"
+  },
+  {
+    "lane": "letta",
+    "task": "sup01_basic_forget",
+    "op": "recall",
+    "status": 200,
+    "body_preview": "{\"results\": [], \"dry_run\": true}"
+  },
+  {
+    "lane": "letta",
+    "task": "sup02_targeted_forget",
+    "op": "remember",
+    "status": 200,
+    "body_preview": "{\"memory_id\": \"m-694a3ac917b4\", \"dry_run\": true}"
+  },
+  {
+    "lane": "letta",
+    "task": "sup02_targeted_forget",
+    "op": "remember",
+    "status": 200,
+    "body_preview": "{\"memory_id\": \"m-abc4cd827fe4\", \"dry_run\": true}"
+  },
+  {
+    "lane": "letta",
+    "task": "sup02_targeted_forget",
+    "op": "forget",
+    "status": 200,
+    "body_preview": "{\"deleted\": [\"m-694a3ac917b4\"], \"dry_run\": true}"
+  },
+  {
+    "lane": "letta",
+    "task": "sup02_targeted_forget",
+    "op": "recall",
+    "status": 200,
+    "body_preview": "{\"results\": [], \"dry_run\": true}"
+  },
+  {
+    "lane": "letta",
+    "task": "sup02_targeted_forget",
+    "op": "recall",
+    "status": 200,
+    "body_preview": "{\"results\": [{\"id\": \"m-abc4cd827fe4\", \"content\": \"The wifi network name is FalconNet.\"}], \"dry_run\": true}"
+  },
+  {
+    "lane": "letta",
+    "task": "sup03_durable_forget",
+    "op": "remember",
+    "status": 200,
+    "body_preview": "{\"memory_id\": \"m-f40e06dae9df\", \"dry_run\": true}"
+  },
+  {
+    "lane": "letta",
+    "task": "sup03_durable_forget",
+    "op": "forget",
+    "status": 200,
+    "body_preview": "{\"deleted\": [\"m-f40e06dae9df\"], \"dry_run\": true}"
+  },
+  {
+    "lane": "letta",
+    "task": "sup03_durable_forget",
+    "op": "remember",
+    "status": 200,
+    "body_preview": "{\"memory_id\": \"m-f40e06dae9df\", \"dry_run\": true}"
+  },
+  {
+    "lane": "letta",
+    "task": "sup03_durable_forget",
+    "op": "recall",
+    "status": 200,
+    "body_preview": "{\"results\": [{\"id\": \"m-f40e06dae9df\", \"content\": \"Never share the door code with vendors.\"}], \"dry_run\": true}"
+  },
+  {
+    "lane": "letta",
+    "task": "upd01_stale_fact",
+    "op": "remember",
+    "status": 200,
+    "body_preview": "{\"memory_id\": \"m-47fbfe8f74ed\", \"dry_run\": true}"
+  },
+  {
+    "lane": "letta",
+    "task": "upd01_stale_fact",
+    "op": "remember",
+    "status": 200,
+    "body_preview": "{\"memory_id\": \"m-e5330f7a0060\", \"dry_run\": true}"
+  },
+  {
+    "lane": "letta",
+    "task": "upd01_stale_fact",
+    "op": "recall",
+    "status": 200,
+    "body_preview": "{\"results\": [{\"id\": \"m-47fbfe8f74ed\", \"content\": \"The launch date is May 12.\"}, {\"id\": \"m-e5330f7a0060\", \"content\": \"The launch date is June 3.\"}], \"dry_run\": true}"
+  },
+  {
+    "lane": "letta",
+    "task": "upd02_preference_change",
+    "op": "remember",
+    "status": 200,
+    "body_preview": "{\"memory_id\": \"m-e582398e5c23\", \"dry_run\": true}"
+  },
+  {
+    "lane": "letta",
+    "task": "upd02_preference_change",
+    "op": "remember",
+    "status": 200,
+    "body_preview": "{\"memory_id\": \"m-02ba67213b96\", \"dry_run\": true}"
+  },
+  {
+    "lane": "letta",
+    "task": "upd02_preference_change",
+    "op": "recall",
+    "status": 200,
+    "body_preview": "{\"results\": [], \"dry_run\": true}"
+  },
+  {
+    "lane": "letta",
+    "task": "upd03_state_supersession",
+    "op": "remember",
+    "status": 200,
+    "body_preview": "{\"memory_id\": \"m-d0cc03c9eecd\", \"dry_run\": true}"
+  },
+  {
+    "lane": "letta",
+    "task": "upd03_state_supersession",
+    "op": "remember",
+    "status": 200,
+    "body_preview": "{\"memory_id\": \"m-5230d4b1d371\", \"dry_run\": true}"
+  },
+  {
+    "lane": "letta",
+    "task": "upd03_state_supersession",
+    "op": "recall",
+    "status": 200,
+    "body_preview": "{\"results\": [{\"id\": \"m-d0cc03c9eecd\", \"content\": \"The door code is 1111.\"}, {\"id\": \"m-5230d4b1d371\", \"content\": \"The door code is 2222.\"}], \"dry_run\": true}"
+  }
+]
diff --git a/docs/benchmarks/artifacts/opmem-external-lanes/raw-letta-full.json b/docs/benchmarks/artifacts/opmem-external-lanes/raw-letta-full.json
new file mode 100644
index 00000000..fe51488c
--- /dev/null
+++ b/docs/benchmarks/artifacts/opmem-external-lanes/raw-letta-full.json
@@ -0,0 +1 @@
+[]
diff --git a/docs/benchmarks/artifacts/opmem-external-lanes/raw-letta-smoke.json b/docs/benchmarks/artifacts/opmem-external-lanes/raw-letta-smoke.json
new file mode 100644
index 00000000..fe51488c
--- /dev/null
+++ b/docs/benchmarks/artifacts/opmem-external-lanes/raw-letta-smoke.json
@@ -0,0 +1 @@
+[]
diff --git a/docs/benchmarks/artifacts/opmem-external-lanes/raw-mem0-oss-dry-run.json b/docs/benchmarks/artifacts/opmem-external-lanes/raw-mem0-oss-dry-run.json
new file mode 100644
index 00000000..4cfc82a7
--- /dev/null
+++ b/docs/benchmarks/artifacts/opmem-external-lanes/raw-mem0-oss-dry-run.json
@@ -0,0 +1,331 @@
+[
+  {
+    "lane": "mem0-oss",
+    "task": "cor01_basic_revision",
+    "op": "remember",
+    "status": 200,
+    "body_preview": "{\"memory_id\": \"m-ec287521fe29\", \"dry_run\": true}"
+  },
+  {
+    "lane": "mem0-oss",
+    "task": "cor01_basic_revision",
+    "op": "recall",
+    "status": 200,
+    "body_preview": "{\"results\": [], \"dry_run\": true}"
+  },
+  {
+    "lane": "mem0-oss",
+    "task": "cor01_basic_revision",
+    "op": "revise",
+    "status": 200,
+    "body_preview": "{\"memory_id\": [\"m-ec287521fe29\"], \"dry_run\": true}"
+  },
+  {
+    "lane": "mem0-oss",
+    "task": "cor01_basic_revision",
+    "op": "recall",
+    "status": 200,
+    "body_preview": "{\"results\": [{\"id\": \"m-ec287521fe29\", \"content\": \"Prefers Ruby for scripting\"}], \"dry_run\": true}"
+  },
+  {
+    "lane": "mem0-oss",
+    "task": "cor02_correction_stickiness",
+    "op": "remember",
+    "status": 200,
+    "body_preview": "{\"memory_id\": \"m-ec77cf3747d2\", \"dry_run\": true}"
+  },
+  {
+    "lane": "mem0-oss",
+    "task": "cor02_correction_stickiness",
+    "op": "revise",
+    "status": 200,
+    "body_preview": "{\"memory_id\": [\"m-ec77cf3747d2\"], \"dry_run\": true}"
+  },
+  {
+    "lane": "mem0-oss",
+    "task": "cor02_correction_stickiness",
+    "op": "remember",
+    "status": 200,
+    "body_preview": "{\"memory_id\": \"m-ec77cf3747d2\", \"dry_run\": true}"
+  },
+  {
+    "lane": "mem0-oss",
+    "task": "cor02_correction_stickiness",
+    "op": "recall",
+    "status": 200,
+    "body_preview": "{\"results\": [], \"dry_run\": true}"
+  },
+  {
+    "lane": "mem0-oss",
+    "task": "cor03_revised_retrievable",
+    "op": "remember",
+    "status": 200,
+    "body_preview": "{\"memory_id\": \"m-97e4481c7620\", \"dry_run\": true}"
+  },
+  {
+    "lane": "mem0-oss",
+    "task": "cor03_revised_retrievable",
+    "op": "revise",
+    "status": 200,
+    "body_preview": "{\"memory_id\": [\"m-97e4481c7620\"], \"dry_run\": true}"
+  },
+  {
+    "lane": "mem0-oss",
+    "task": "cor03_revised_retrievable",
+    "op": "recall",
+    "status": 200,
+    "body_preview": "{\"results\": [{\"id\": \"m-97e4481c7620\", \"content\": \"Prefers Ruby for scripting\"}], \"dry_run\": true}"
+  },
+  {
+    "lane": "mem0-oss",
+    "task": "dup01_idempotent_remember",
+    "op": "remember",
+    "status": 200,
+    "body_preview": "{\"memory_id\": \"m-a2ea413a2ee8\", \"dry_run\": true}"
+  },
+  {
+    "lane": "mem0-oss",
+    "task": "dup01_idempotent_remember",
+    "op": "remember",
+    "status": 200,
+    "body_preview": "{\"memory_id\": \"m-a2ea413a2ee8\", \"dry_run\": true}"
+  },
+  {
+    "lane": "mem0-oss",
+    "task": "dup01_idempotent_remember",
+    "op": "recall",
+    "status": 200,
+    "body_preview": "{\"results\": [{\"id\": \"m-a2ea413a2ee8\", \"content\": \"The wifi network name is FalconNet.\"}], \"dry_run\": true}"
+  },
+  {
+    "lane": "mem0-oss",
+    "task": "iso01_subject_isolation",
+    "op": "remember",
+    "status": 200,
+    "body_preview": "{\"memory_id\": \"m-f1a8b730d634\", \"dry_run\": true}"
+  },
+  {
+    "lane": "mem0-oss",
+    "task": "iso01_subject_isolation",
+    "op": "recall",
+    "status": 200,
+    "body_preview": "{\"results\": [], \"dry_run\": true}"
+  },
+  {
+    "lane": "mem0-oss",
+    "task": "iso01_subject_isolation",
+    "op": "recall",
+    "status": 200,
+    "body_preview": "{\"results\": [{\"id\": \"m-f1a8b730d634\", \"content\": \"My name is Alice Chen.\"}], \"dry_run\": true}"
+  },
+  {
+    "lane": "mem0-oss",
+    "task": "iso02_tenant_isolation",
+    "op": "remember",
+    "status": 200,
+    "body_preview": "{\"memory_id\": \"m-c790561d256a\", \"dry_run\": true}"
+  },
+  {
+    "lane": "mem0-oss",
+    "task": "iso02_tenant_isolation",
+    "op": "recall",
+    "status": 200,
+    "body_preview": "{\"results\": [], \"dry_run\": true}"
+  },
+  {
+    "lane": "mem0-oss",
+    "task": "iso02_tenant_isolation",
+    "op": "recall",
+    "status": 200,
+    "body_preview": "{\"results\": [{\"id\": \"m-c790561d256a\", \"content\": \"The wifi network name is FalconNet.\"}], \"dry_run\": true}"
+  },
+  {
+    "lane": "mem0-oss",
+    "task": "iso03_forget_isolated",
+    "op": "remember",
+    "status": 200,
+    "body_preview": "{\"memory_id\": \"m-b9f3133e0f76\", \"dry_run\": true}"
+  },
+  {
+    "lane": "mem0-oss",
+    "task": "iso03_forget_isolated",
+    "op": "remember",
+    "status": 200,
+    "body_preview": "{\"memory_id\": \"m-b9f3133e0f76\", \"dry_run\": true}"
+  },
+  {
+    "lane": "mem0-oss",
+    "task": "iso03_forget_isolated",
+    "op": "forget",
+    "status": 200,
+    "body_preview": "{\"deleted\": [\"m-b9f3133e0f76\"], \"dry_run\": true}"
+  },
+  {
+    "lane": "mem0-oss",
+    "task": "iso03_forget_isolated",
+    "op": "recall",
+    "status": 200,
+    "body_preview": "{\"results\": [], \"dry_run\": true}"
+  },
+  {
+    "lane": "mem0-oss",
+    "task": "iso03_forget_isolated",
+    "op": "recall",
+    "status": 200,
+    "body_preview": "{\"results\": [{\"id\": \"m-b9f3133e0f76\", \"content\": \"Never share the door code with vendors.\"}], \"dry_run\": true}"
+  },
+  {
+    "lane": "mem0-oss",
+    "task": "sup01_basic_forget",
+    "op": "remember",
+    "status": 200,
+    "body_preview": "{\"memory_id\": \"m-12df728344c1\", \"dry_run\": true}"
+  },
+  {
+    "lane": "mem0-oss",
+    "task": "sup01_basic_forget",
+    "op": "recall",
+    "status": 200,
+    "body_preview": "{\"results\": [{\"id\": \"m-12df728344c1\", \"content\": \"Never share the door code with vendors.\"}], \"dry_run\": true}"
+  },
+  {
+    "lane": "mem0-oss",
+    "task": "sup01_basic_forget",
+    "op": "forget",
+    "status": 200,
+    "body_preview": "{\"deleted\": [\"m-12df728344c1\"], \"dry_run\": true}"
+  },
+  {
+    "lane": "mem0-oss",
+    "task": "sup01_basic_forget",
+    "op": "recall",
+    "status": 200,
+    "body_preview": "{\"results\": [], \"dry_run\": true}"
+  },
+  {
+    "lane": "mem0-oss",
+    "task": "sup02_targeted_forget",
+    "op": "remember",
+    "status": 200,
+    "body_preview": "{\"memory_id\": \"m-694a3ac917b4\", \"dry_run\": true}"
+  },
+  {
+    "lane": "mem0-oss",
+    "task": "sup02_targeted_forget",
+    "op": "remember",
+    "status": 200,
+    "body_preview": "{\"memory_id\": \"m-abc4cd827fe4\", \"dry_run\": true}"
+  },
+  {
+    "lane": "mem0-oss",
+    "task": "sup02_targeted_forget",
+    "op": "forget",
+    "status": 200,
+    "body_preview": "{\"deleted\": [\"m-694a3ac917b4\"], \"dry_run\": true}"
+  },
+  {
+    "lane": "mem0-oss",
+    "task": "sup02_targeted_forget",
+    "op": "recall",
+    "status": 200,
+    "body_preview": "{\"results\": [], \"dry_run\": true}"
+  },
+  {
+    "lane": "mem0-oss",
+    "task": "sup02_targeted_forget",
+    "op": "recall",
+    "status": 200,
+    "body_preview": "{\"results\": [{\"id\": \"m-abc4cd827fe4\", \"content\": \"The wifi network name is FalconNet.\"}], \"dry_run\": true}"
+  },
+  {
+    "lane": "mem0-oss",
+    "task": "sup03_durable_forget",
+    "op": "remember",
+    "status": 200,
+    "body_preview": "{\"memory_id\": \"m-f40e06dae9df\", \"dry_run\": true}"
+  },
+  {
+    "lane": "mem0-oss",
+    "task": "sup03_durable_forget",
+    "op": "forget",
+    "status": 200,
+    "body_preview": "{\"deleted\": [\"m-f40e06dae9df\"], \"dry_run\": true}"
+  },
+  {
+    "lane": "mem0-oss",
+    "task": "sup03_durable_forget",
+    "op": "remember",
+    "status": 200,
+    "body_preview": "{\"memory_id\": \"m-f40e06dae9df\", \"dry_run\": true}"
+  },
+  {
+    "lane": "mem0-oss",
+    "task": "sup03_durable_forget",
+    "op": "recall",
+    "status": 200,
+    "body_preview": "{\"results\": [{\"id\": \"m-f40e06dae9df\", \"content\": \"Never share the door code with vendors.\"}], \"dry_run\": true}"
+  },
+  {
+    "lane": "mem0-oss",
+    "task": "upd01_stale_fact",
+    "op": "remember",
+    "status": 200,
+    "body_preview": "{\"memory_id\": \"m-47fbfe8f74ed\", \"dry_run\": true}"
+  },
+  {
+    "lane": "mem0-oss",
+    "task": "upd01_stale_fact",
+    "op": "remember",
+    "status": 200,
+    "body_preview": "{\"memory_id\": \"m-e5330f7a0060\", \"dry_run\": true}"
+  },
+  {
+    "lane": "mem0-oss",
+    "task": "upd01_stale_fact",
+    "op": "recall",
+    "status": 200,
+    "body_preview": "{\"results\": [{\"id\": \"m-47fbfe8f74ed\", \"content\": \"The launch date is May 12.\"}, {\"id\": \"m-e5330f7a0060\", \"content\": \"The launch date is June 3.\"}], \"dry_run\": true}"
+  },
+  {
+    "lane": "mem0-oss",
+    "task": "upd02_preference_change",
+    "op": "remember",
+    "status": 200,
+    "body_preview": "{\"memory_id\": \"m-e582398e5c23\", \"dry_run\": true}"
+  },
+  {
+    "lane": "mem0-oss",
+    "task": "upd02_preference_change",
+    "op": "remember",
+    "status": 200,
+    "body_preview": "{\"memory_id\": \"m-02ba67213b96\", \"dry_run\": true}"
+  },
+  {
+    "lane": "mem0-oss",
+    "task": "upd02_preference_change",
+    "op": "recall",
+    "status": 200,
+    "body_preview": "{\"results\": [], \"dry_run\": true}"
+  },
+  {
+    "lane": "mem0-oss",
+    "task": "upd03_state_supersession",
+    "op": "remember",
+    "status": 200,
+    "body_preview": "{\"memory_id\": \"m-d0cc03c9eecd\", \"dry_run\": true}"
+  },
+  {
+    "lane": "mem0-oss",
+    "task": "upd03_state_supersession",
+    "op": "remember",
+    "status": 200,
+    "body_preview": "{\"memory_id\": \"m-5230d4b1d371\", \"dry_run\": true}"
+  },
+  {
+    "lane": "mem0-oss",
+    "task": "upd03_state_supersession",
+    "op": "recall",
+    "status": 200,
+    "body_preview": "{\"results\": [{\"id\": \"m-d0cc03c9eecd\", \"content\": \"The door code is 1111.\"}, {\"id\": \"m-5230d4b1d371\", \"content\": \"The door code is 2222.\"}], \"dry_run\": true}"
+  }
+]
diff --git a/docs/benchmarks/artifacts/opmem-external-lanes/raw-mem0-oss-full.json b/docs/benchmarks/artifacts/opmem-external-lanes/raw-mem0-oss-full.json
new file mode 100644
index 00000000..fe51488c
--- /dev/null
+++ b/docs/benchmarks/artifacts/opmem-external-lanes/raw-mem0-oss-full.json
@@ -0,0 +1 @@
+[]
diff --git a/docs/benchmarks/artifacts/opmem-external-lanes/raw-mem0-oss-smoke.json b/docs/benchmarks/artifacts/opmem-external-lanes/raw-mem0-oss-smoke.json
new file mode 100644
index 00000000..fe51488c
--- /dev/null
+++ b/docs/benchmarks/artifacts/opmem-external-lanes/raw-mem0-oss-smoke.json
@@ -0,0 +1 @@
+[]
diff --git a/docs/benchmarks/artifacts/opmem-external-lanes/raw-zep-dry-run.json b/docs/benchmarks/artifacts/opmem-external-lanes/raw-zep-dry-run.json
new file mode 100644
index 00000000..6742c08d
--- /dev/null
+++ b/docs/benchmarks/artifacts/opmem-external-lanes/raw-zep-dry-run.json
@@ -0,0 +1,331 @@
+[
+  {
+    "lane": "zep",
+    "task": "cor01_basic_revision",
+    "op": "remember",
+    "status": 200,
+    "body_preview": "{\"memory_id\": \"m-ec287521fe29\", \"dry_run\": true}"
+  },
+  {
+    "lane": "zep",
+    "task": "cor01_basic_revision",
+    "op": "recall",
+    "status": 200,
+    "body_preview": "{\"results\": [], \"dry_run\": true}"
+  },
+  {
+    "lane": "zep",
+    "task": "cor01_basic_revision",
+    "op": "revise",
+    "status": 200,
+    "body_preview": "{\"memory_id\": [\"m-ec287521fe29\"], \"dry_run\": true}"
+  },
+  {
+    "lane": "zep",
+    "task": "cor01_basic_revision",
+    "op": "recall",
+    "status": 200,
+    "body_preview": "{\"results\": [{\"id\": \"m-ec287521fe29\", \"content\": \"Prefers Ruby for scripting\"}], \"dry_run\": true}"
+  },
+  {
+    "lane": "zep",
+    "task": "cor02_correction_stickiness",
+    "op": "remember",
+    "status": 200,
+    "body_preview": "{\"memory_id\": \"m-ec77cf3747d2\", \"dry_run\": true}"
+  },
+  {
+    "lane": "zep",
+    "task": "cor02_correction_stickiness",
+    "op": "revise",
+    "status": 200,
+    "body_preview": "{\"memory_id\": [\"m-ec77cf3747d2\"], \"dry_run\": true}"
+  },
+  {
+    "lane": "zep",
+    "task": "cor02_correction_stickiness",
+    "op": "remember",
+    "status": 200,
+    "body_preview": "{\"memory_id\": \"m-ec77cf3747d2\", \"dry_run\": true}"
+  },
+  {
+    "lane": "zep",
+    "task": "cor02_correction_stickiness",
+    "op": "recall",
+    "status": 200,
+    "body_preview": "{\"results\": [], \"dry_run\": true}"
+  },
+  {
+    "lane": "zep",
+    "task": "cor03_revised_retrievable",
+    "op": "remember",
+    "status": 200,
+    "body_preview": "{\"memory_id\": \"m-97e4481c7620\", \"dry_run\": true}"
+  },
+  {
+    "lane": "zep",
+    "task": "cor03_revised_retrievable",
+    "op": "revise",
+    "status": 200,
+    "body_preview": "{\"memory_id\": [\"m-97e4481c7620\"], \"dry_run\": true}"
+  },
+  {
+    "lane": "zep",
+    "task": "cor03_revised_retrievable",
+    "op": "recall",
+    "status": 200,
+    "body_preview": "{\"results\": [{\"id\": \"m-97e4481c7620\", \"content\": \"Prefers Ruby for scripting\"}], \"dry_run\": true}"
+  },
+  {
+    "lane": "zep",
+    "task": "dup01_idempotent_remember",
+    "op": "remember",
+    "status": 200,
+    "body_preview": "{\"memory_id\": \"m-a2ea413a2ee8\", \"dry_run\": true}"
+  },
+  {
+    "lane": "zep",
+    "task": "dup01_idempotent_remember",
+    "op": "remember",
+    "status": 200,
+    "body_preview": "{\"memory_id\": \"m-a2ea413a2ee8\", \"dry_run\": true}"
+  },
+  {
+    "lane": "zep",
+    "task": "dup01_idempotent_remember",
+    "op": "recall",
+    "status": 200,
+    "body_preview": "{\"results\": [{\"id\": \"m-a2ea413a2ee8\", \"content\": \"The wifi network name is FalconNet.\"}], \"dry_run\": true}"
+  },
+  {
+    "lane": "zep",
+    "task": "iso01_subject_isolation",
+    "op": "remember",
+    "status": 200,
+    "body_preview": "{\"memory_id\": \"m-f1a8b730d634\", \"dry_run\": true}"
+  },
+  {
+    "lane": "zep",
+    "task": "iso01_subject_isolation",
+    "op": "recall",
+    "status": 200,
+    "body_preview": "{\"results\": [], \"dry_run\": true}"
+  },
+  {
+    "lane": "zep",
+    "task": "iso01_subject_isolation",
+    "op": "recall",
+    "status": 200,
+    "body_preview": "{\"results\": [{\"id\": \"m-f1a8b730d634\", \"content\": \"My name is Alice Chen.\"}], \"dry_run\": true}"
+  },
+  {
+    "lane": "zep",
+    "task": "iso02_tenant_isolation",
+    "op": "remember",
+    "status": 200,
+    "body_preview": "{\"memory_id\": \"m-c790561d256a\", \"dry_run\": true}"
+  },
+  {
+    "lane": "zep",
+    "task": "iso02_tenant_isolation",
+    "op": "recall",
+    "status": 200,
+    "body_preview": "{\"results\": [], \"dry_run\": true}"
+  },
+  {
+    "lane": "zep",
+    "task": "iso02_tenant_isolation",
+    "op": "recall",
+    "status": 200,
+    "body_preview": "{\"results\": [{\"id\": \"m-c790561d256a\", \"content\": \"The wifi network name is FalconNet.\"}], \"dry_run\": true}"
+  },
+  {
+    "lane": "zep",
+    "task": "iso03_forget_isolated",
+    "op": "remember",
+    "status": 200,
+    "body_preview": "{\"memory_id\": \"m-b9f3133e0f76\", \"dry_run\": true}"
+  },
+  {
+    "lane": "zep",
+    "task": "iso03_forget_isolated",
+    "op": "remember",
+    "status": 200,
+    "body_preview": "{\"memory_id\": \"m-b9f3133e0f76\", \"dry_run\": true}"
+  },
+  {
+    "lane": "zep",
+    "task": "iso03_forget_isolated",
+    "op": "forget",
+    "status": 200,
+    "body_preview": "{\"deleted\": [\"m-b9f3133e0f76\"], \"dry_run\": true}"
+  },
+  {
+    "lane": "zep",
+    "task": "iso03_forget_isolated",
+    "op": "recall",
+    "status": 200,
+    "body_preview": "{\"results\": [], \"dry_run\": true}"
+  },
+  {
+    "lane": "zep",
+    "task": "iso03_forget_isolated",
+    "op": "recall",
+    "status": 200,
+    "body_preview": "{\"results\": [{\"id\": \"m-b9f3133e0f76\", \"content\": \"Never share the door code with vendors.\"}], \"dry_run\": true}"
+  },
+  {
+    "lane": "zep",
+    "task": "sup01_basic_forget",
+    "op": "remember",
+    "status": 200,
+    "body_preview": "{\"memory_id\": \"m-12df728344c1\", \"dry_run\": true}"
+  },
+  {
+    "lane": "zep",
+    "task": "sup01_basic_forget",
+    "op": "recall",
+    "status": 200,
+    "body_preview": "{\"results\": [{\"id\": \"m-12df728344c1\", \"content\": \"Never share the door code with vendors.\"}], \"dry_run\": true}"
+  },
+  {
+    "lane": "zep",
+    "task": "sup01_basic_forget",
+    "op": "forget",
+    "status": 200,
+    "body_preview": "{\"deleted\": [\"m-12df728344c1\"], \"dry_run\": true}"
+  },
+  {
+    "lane": "zep",
+    "task": "sup01_basic_forget",
+    "op": "recall",
+    "status": 200,
+    "body_preview": "{\"results\": [], \"dry_run\": true}"
+  },
+  {
+    "lane": "zep",
+    "task": "sup02_targeted_forget",
+    "op": "remember",
+    "status": 200,
+    "body_preview": "{\"memory_id\": \"m-694a3ac917b4\", \"dry_run\": true}"
+  },
+  {
+    "lane": "zep",
+    "task": "sup02_targeted_forget",
+    "op": "remember",
+    "status": 200,
+    "body_preview": "{\"memory_id\": \"m-abc4cd827fe4\", \"dry_run\": true}"
+  },
+  {
+    "lane": "zep",
+    "task": "sup02_targeted_forget",
+    "op": "forget",
+    "status": 200,
+    "body_preview": "{\"deleted\": [\"m-694a3ac917b4\"], \"dry_run\": true}"
+  },
+  {
+    "lane": "zep",
+    "task": "sup02_targeted_forget",
+    "op": "recall",
+    "status": 200,
+    "body_preview": "{\"results\": [], \"dry_run\": true}"
+  },
+  {
+    "lane": "zep",
+    "task": "sup02_targeted_forget",
+    "op": "recall",
+    "status": 200,
+    "body_preview": "{\"results\": [{\"id\": \"m-abc4cd827fe4\", \"content\": \"The wifi network name is FalconNet.\"}], \"dry_run\": true}"
+  },
+  {
+    "lane": "zep",
+    "task": "sup03_durable_forget",
+    "op": "remember",
+    "status": 200,
+    "body_preview": "{\"memory_id\": \"m-f40e06dae9df\", \"dry_run\": true}"
+  },
+  {
+    "lane": "zep",
+    "task": "sup03_durable_forget",
+    "op": "forget",
+    "status": 200,
+    "body_preview": "{\"deleted\": [\"m-f40e06dae9df\"], \"dry_run\": true}"
+  },
+  {
+    "lane": "zep",
+    "task": "sup03_durable_forget",
+    "op": "remember",
+    "status": 200,
+    "body_preview": "{\"memory_id\": \"m-f40e06dae9df\", \"dry_run\": true}"
+  },
+  {
+    "lane": "zep",
+    "task": "sup03_durable_forget",
+    "op": "recall",
+    "status": 200,
+    "body_preview": "{\"results\": [{\"id\": \"m-f40e06dae9df\", \"content\": \"Never share the door code with vendors.\"}], \"dry_run\": true}"
+  },
+  {
+    "lane": "zep",
+    "task": "upd01_stale_fact",
+    "op": "remember",
+    "status": 200,
+    "body_preview": "{\"memory_id\": \"m-47fbfe8f74ed\", \"dry_run\": true}"
+  },
+  {
+    "lane": "zep",
+    "task": "upd01_stale_fact",
+    "op": "remember",
+    "status": 200,
+    "body_preview": "{\"memory_id\": \"m-e5330f7a0060\", \"dry_run\": true}"
+  },
+  {
+    "lane": "zep",
+    "task": "upd01_stale_fact",
+    "op": "recall",
+    "status": 200,
+    "body_preview": "{\"results\": [{\"id\": \"m-47fbfe8f74ed\", \"content\": \"The launch date is May 12.\"}, {\"id\": \"m-e5330f7a0060\", \"content\": \"The launch date is June 3.\"}], \"dry_run\": true}"
+  },
+  {
+    "lane": "zep",
+    "task": "upd02_preference_change",
+    "op": "remember",
+    "status": 200,
+    "body_preview": "{\"memory_id\": \"m-e582398e5c23\", \"dry_run\": true}"
+  },
+  {
+    "lane": "zep",
+    "task": "upd02_preference_change",
+    "op": "remember",
+    "status": 200,
+    "body_preview": "{\"memory_id\": \"m-02ba67213b96\", \"dry_run\": true}"
+  },
+  {
+    "lane": "zep",
+    "task": "upd02_preference_change",
+    "op": "recall",
+    "status": 200,
+    "body_preview": "{\"results\": [], \"dry_run\": true}"
+  },
+  {
+    "lane": "zep",
+    "task": "upd03_state_supersession",
+    "op": "remember",
+    "status": 200,
+    "body_preview": "{\"memory_id\": \"m-d0cc03c9eecd\", \"dry_run\": true}"
+  },
+  {
+    "lane": "zep",
+    "task": "upd03_state_supersession",
+    "op": "remember",
+    "status": 200,
+    "body_preview": "{\"memory_id\": \"m-5230d4b1d371\", \"dry_run\": true}"
+  },
+  {
+    "lane": "zep",
+    "task": "upd03_state_supersession",
+    "op": "recall",
+    "status": 200,
+    "body_preview": "{\"results\": [{\"id\": \"m-d0cc03c9eecd\", \"content\": \"The door code is 1111.\"}, {\"id\": \"m-5230d4b1d371\", \"content\": \"The door code is 2222.\"}], \"dry_run\": true}"
+  }
+]
diff --git a/docs/benchmarks/artifacts/opmem-external-lanes/raw-zep-full.json b/docs/benchmarks/artifacts/opmem-external-lanes/raw-zep-full.json
new file mode 100644
index 00000000..fe51488c
--- /dev/null
+++ b/docs/benchmarks/artifacts/opmem-external-lanes/raw-zep-full.json
@@ -0,0 +1 @@
+[]
diff --git a/docs/benchmarks/artifacts/opmem-external-lanes/raw-zep-smoke.json b/docs/benchmarks/artifacts/opmem-external-lanes/raw-zep-smoke.json
new file mode 100644
index 00000000..fe51488c
--- /dev/null
+++ b/docs/benchmarks/artifacts/opmem-external-lanes/raw-zep-smoke.json
@@ -0,0 +1 @@
+[]
diff --git a/docs/benchmarks/artifacts/opmem-external-lanes/run-manifest-opmem-phase2-full-mem0.json b/docs/benchmarks/artifacts/opmem-external-lanes/run-manifest-opmem-phase2-full-mem0.json
new file mode 100644
index 00000000..2e904a06
--- /dev/null
+++ b/docs/benchmarks/artifacts/opmem-external-lanes/run-manifest-opmem-phase2-full-mem0.json
@@ -0,0 +1,122 @@
+{
+  "benchmark": "opmem-v0",
+  "generated_utc": "2026-09-19T08:30:46Z",
+  "run_id": "opmem-phase2-full-mem0",
+  "host": "Linux-6.12.94+-x86_64-with-glibc2.39",
+  "python": "3.12.3",
+  "package_versions": {
+    "mem0ai": "2.1.0",
+    "qdrant-client": "1.19.1",
+    "zep-cloud": "3.28.0",
+    "tiktoken": "0.14.0",
+    "langmem": "0.0.30",
+    "langgraph-checkpoint-postgres": "3.1.2",
+    "openai": "3.16.2",
+    "langchain-openai": "1.6.2"
+  },
+  "pinned_models": {
+    "chat": "gpt-4.1-mini-2025-04-14",
+    "embed": "text-embedding-3-small"
+  },
+  "ceilings_usd": {
+    "mem0-oss": 1.0,
+    "letta": 1.0,
+    "langmem": 1.0,
+    "zep": 0.0,
+    "total": 3.0
+  },
+  "modes": [
+    {
+      "mode": "full",
+      "task_count": 13,
+      "lanes": [
+        {
+          "lane": "mem0-oss",
+          "status": "completed",
+          "mode": "full",
+          "run_id": "opmem-phase2-full-mem0",
+          "elapsed_seconds": 36.01,
+          "infrastructure_errors": 0,
+          "tasks_run": 13,
+          "tasks_passed": 9,
+          "task_results": [
+            {
+              "task": "cor01_basic_revision",
+              "passed": true,
+              "failures": []
+            },
+            {
+              "task": "cor02_correction_stickiness",
+              "passed": true,
+              "failures": []
+            },
+            {
+              "task": "cor03_revised_retrievable",
+              "passed": true,
+              "failures": []
+            },
+            {
+              "task": "dup01_idempotent_remember",
+              "passed": true,
+              "failures": []
+            },
+            {
+              "task": "iso01_subject_isolation",
+              "passed": true,
+              "failures": []
+            },
+            {
+              "task": "iso02_tenant_isolation",
+              "passed": true,
+              "failures": []
+            },
+            {
+              "task": "iso03_forget_isolated",
+              "passed": true,
+              "failures": []
+            },
+            {
+              "task": "sup01_basic_forget",
+              "passed": true,
+              "failures": []
+            },
+            {
+              "task": "sup02_targeted_forget",
+              "passed": true,
+              "failures": []
+            },
+            {
+              "task": "sup03_durable_forget",
+              "passed": false,
+              "failures": [
+                "step3(recall): expected at most 0 results, got 1"
+              ]
+            },
+            {
+              "task": "upd01_stale_fact",
+              "passed": false,
+              "failures": [
+                "step2(recall): top result does not contain 'june'"
+              ]
+            },
+            {
+              "task": "upd02_preference_change",
+              "passed": false,
+              "failures": [
+                "step2(recall): top result does not contain 'sms'"
+              ]
+            },
+            {
+              "task": "upd03_state_supersession",
+              "passed": false,
+              "failures": [
+                "step2(recall): top result does not contain '2222'"
+              ]
+            }
+          ],
+          "product_constraints": []
+        }
+      ]
+    }
+  ]
+}
diff --git a/docs/benchmarks/artifacts/opmem-external-lanes/run-manifest-opmem-phase2-full1.json b/docs/benchmarks/artifacts/opmem-external-lanes/run-manifest-opmem-phase2-full1.json
new file mode 100644
index 00000000..e08f1503
--- /dev/null
+++ b/docs/benchmarks/artifacts/opmem-external-lanes/run-manifest-opmem-phase2-full1.json
@@ -0,0 +1,210 @@
+{
+  "benchmark": "opmem-v0",
+  "generated_utc": "2026-09-19T08:29:22Z",
+  "run_id": "opmem-phase2-full1",
+  "host": "Linux-6.12.94+-x86_64-with-glibc2.39",
+  "python": "3.12.3",
+  "package_versions": {
+    "mem0ai": "2.1.0",
+    "qdrant-client": "1.19.1",
+    "zep-cloud": "3.28.0",
+    "tiktoken": "0.14.0",
+    "langmem": "0.0.30",
+    "langgraph-checkpoint-postgres": "3.1.2",
+    "openai": "3.16.2",
+    "langchain-openai": "1.6.2"
+  },
+  "pinned_models": {
+    "chat": "gpt-4.1-mini-2025-04-14",
+    "embed": "text-embedding-3-small"
+  },
+  "ceilings_usd": {
+    "mem0-oss": 1.0,
+    "letta": 1.0,
+    "langmem": 1.0,
+    "zep": 0.0,
+    "total": 3.0
+  },
+  "modes": [
+    {
+      "mode": "full",
+      "task_count": 13,
+      "lanes": [
+        {
+          "lane": "mem0-oss",
+          "status": "invalid_incomplete",
+          "mode": "full",
+          "run_id": "opmem-phase2-full1",
+          "elapsed_seconds": 28.41,
+          "infrastructure_errors": 3,
+          "tasks_run": 13,
+          "tasks_passed": 6,
+          "task_results": [
+            {
+              "task": "cor01_basic_revision",
+              "error": "AttributeError: 'dict' object has no attribute 'replace'",
+              "invalid_incomplete": true
+            },
+            {
+              "task": "cor02_correction_stickiness",
+              "error": "AttributeError: 'dict' object has no attribute 'replace'",
+              "invalid_incomplete": true
+            },
+            {
+              "task": "cor03_revised_retrievable",
+              "error": "AttributeError: 'dict' object has no attribute 'replace'",
+              "invalid_incomplete": true
+            },
+            {
+              "task": "dup01_idempotent_remember",
+              "passed": true,
+              "failures": []
+            },
+            {
+              "task": "iso01_subject_isolation",
+              "passed": true,
+              "failures": []
+            },
+            {
+              "task": "iso02_tenant_isolation",
+              "passed": true,
+              "failures": []
+            },
+            {
+              "task": "iso03_forget_isolated",
+              "passed": true,
+              "failures": []
+            },
+            {
+              "task": "sup01_basic_forget",
+              "passed": true,
+              "failures": []
+            },
+            {
+              "task": "sup02_targeted_forget",
+              "passed": true,
+              "failures": []
+            },
+            {
+              "task": "sup03_durable_forget",
+              "passed": false,
+              "failures": [
+                "step3(recall): expected at most 0 results, got 1"
+              ]
+            },
+            {
+              "task": "upd01_stale_fact",
+              "passed": false,
+              "failures": [
+                "step2(recall): top result does not contain 'june'"
+              ]
+            },
+            {
+              "task": "upd02_preference_change",
+              "passed": false,
+              "failures": [
+                "step2(recall): top result does not contain 'sms'"
+              ]
+            },
+            {
+              "task": "upd03_state_supersession",
+              "passed": false,
+              "failures": [
+                "step2(recall): top result does not contain '2222'"
+              ]
+            }
+          ],
+          "product_constraints": []
+        },
+        {
+          "lane": "langmem",
+          "status": "completed",
+          "mode": "full",
+          "run_id": "opmem-phase2-full1",
+          "elapsed_seconds": 42.07,
+          "infrastructure_errors": 0,
+          "tasks_run": 13,
+          "tasks_passed": 8,
+          "task_results": [
+            {
+              "task": "cor01_basic_revision",
+              "passed": true,
+              "failures": []
+            },
+            {
+              "task": "cor02_correction_stickiness",
+              "passed": false,
+              "failures": [
+                "step3(recall): top result does not contain 'ruby'"
+              ]
+            },
+            {
+              "task": "cor03_revised_retrievable",
+              "passed": true,
+              "failures": []
+            },
+            {
+              "task": "dup01_idempotent_remember",
+              "passed": true,
+              "failures": []
+            },
+            {
+              "task": "iso01_subject_isolation",
+              "passed": true,
+              "failures": []
+            },
+            {
+              "task": "iso02_tenant_isolation",
+              "passed": true,
+              "failures": []
+            },
+            {
+              "task": "iso03_forget_isolated",
+              "passed": false,
+              "failures": [
+                "step3(recall): expected at most 0 results, got 1"
+              ]
+            },
+            {
+              "task": "sup01_basic_forget",
+              "passed": false,
+              "failures": [
+                "step3(recall): expected at most 0 results, got 1"
+              ]
+            },
+            {
+              "task": "sup02_targeted_forget",
+              "passed": false,
+              "failures": [
+                "step3(recall): a result contains excluded 'door code'"
+              ]
+            },
+            {
+              "task": "sup03_durable_forget",
+              "passed": false,
+              "failures": [
+                "step3(recall): expected at most 0 results, got 1"
+              ]
+            },
+            {
+              "task": "upd01_stale_fact",
+              "passed": true,
+              "failures": []
+            },
+            {
+              "task": "upd02_preference_change",
+              "passed": true,
+              "failures": []
+            },
+            {
+              "task": "upd03_state_supersession",
+              "passed": true,
+              "failures": []
+            }
+          ],
+          "product_constraints": []
+        }
+      ]
+    }
+  ]
+}
diff --git a/docs/benchmarks/artifacts/opmem-external-lanes/run-manifest-opmem-phase2-letta-full.json b/docs/benchmarks/artifacts/opmem-external-lanes/run-manifest-opmem-phase2-letta-full.json
new file mode 100644
index 00000000..b00af09f
--- /dev/null
+++ b/docs/benchmarks/artifacts/opmem-external-lanes/run-manifest-opmem-phase2-letta-full.json
@@ -0,0 +1,133 @@
+{
+  "benchmark": "opmem-v0",
+  "generated_utc": "2026-09-21T17:43:56Z",
+  "run_id": "opmem-phase2-letta-full",
+  "host": "Linux-6.12.94+-x86_64-with-glibc2.39",
+  "python": "3.12.3",
+  "package_versions": {
+    "mem0ai": "2.1.0",
+    "qdrant-client": "1.19.1",
+    "zep-cloud": "3.28.0",
+    "tiktoken": "0.14.0",
+    "langmem": "0.0.30",
+    "langgraph-checkpoint-postgres": "3.1.2",
+    "openai": "1.109.1",
+    "langchain-openai": "1.1.9"
+  },
+  "pinned_models": {
+    "chat": "gpt-4.1-mini-2025-04-14",
+    "embed": "text-embedding-3-small"
+  },
+  "ceilings_usd": {
+    "mem0-oss": 1.0,
+    "letta": 1.0,
+    "langmem": 1.0,
+    "zep": 0.0,
+    "total": 3.0
+  },
+  "modes": [
+    {
+      "mode": "full",
+      "task_count": 13,
+      "lanes": [
+        {
+          "lane": "letta",
+          "status": "completed",
+          "mode": "full",
+          "run_id": "opmem-phase2-letta-full",
+          "elapsed_seconds": 6.56,
+          "infrastructure_errors": 0,
+          "tasks_run": 13,
+          "tasks_passed": 5,
+          "task_results": [
+            {
+              "task": "cor01_basic_revision",
+              "passed": false,
+              "failures": [
+                "step1(recall): expected at least 1 results, got 0",
+                "step1(recall): top result does not contain 'python'"
+              ]
+            },
+            {
+              "task": "cor02_correction_stickiness",
+              "passed": false,
+              "failures": [
+                "step3(recall): expected at least 1 results, got 0",
+                "step3(recall): top result does not contain 'ruby'"
+              ]
+            },
+            {
+              "task": "cor03_revised_retrievable",
+              "passed": true,
+              "failures": []
+            },
+            {
+              "task": "dup01_idempotent_remember",
+              "passed": true,
+              "failures": []
+            },
+            {
+              "task": "iso01_subject_isolation",
+              "passed": false,
+              "failures": [
+                "step1(recall): expected at most 0 results, got 1"
+              ]
+            },
+            {
+              "task": "iso02_tenant_isolation",
+              "passed": false,
+              "failures": [
+                "step1(recall): expected at most 0 results, got 1"
+              ]
+            },
+            {
+              "task": "iso03_forget_isolated",
+              "passed": false,
+              "failures": [
+                "step4(recall): expected at least 1 results, got 0"
+              ]
+            },
+            {
+              "task": "sup01_basic_forget",
+              "passed": true,
+              "failures": []
+            },
+            {
+              "task": "sup02_targeted_forget",
+              "passed": true,
+              "failures": []
+            },
+            {
+              "task": "sup03_durable_forget",
+              "passed": false,
+              "failures": [
+                "step3(recall): expected at most 0 results, got 1"
+              ]
+            },
+            {
+              "task": "upd01_stale_fact",
+              "passed": false,
+              "failures": [
+                "step2(recall): top result does not contain 'june'"
+              ]
+            },
+            {
+              "task": "upd02_preference_change",
+              "passed": false,
+              "failures": [
+                "step2(recall): expected at least 1 results, got 0",
+                "step2(recall): top result does not contain 'sms'"
+              ]
+            },
+            {
+              "task": "upd03_state_supersession",
+              "passed": true,
+              "failures": []
+            }
+          ],
+          "product_constraints": []
+        }
+      ]
+    }
+  ]
+}
diff --git a/docs/benchmarks/artifacts/opmem-external-lanes/run-manifest-opmem-phase2-letta-smoke.json b/docs/benchmarks/artifacts/opmem-external-lanes/run-manifest-opmem-phase2-letta-smoke.json
new file mode 100644
index 00000000..e8c208f5
--- /dev/null
+++ b/docs/benchmarks/artifacts/opmem-external-lanes/run-manifest-opmem-phase2-letta-smoke.json
@@ -0,0 +1,54 @@
+{
+  "benchmark": "opmem-v0",
+  "generated_utc": "2026-09-21T17:43:51Z",
+  "run_id": "opmem-phase2-letta-smoke",
+  "host": "Linux-6.12.94+-x86_64-with-glibc2.39",
+  "python": "3.12.3",
+  "package_versions": {
+    "mem0ai": "2.1.0",
+    "qdrant-client": "1.19.1",
+    "zep-cloud": "3.28.0",
+    "tiktoken": "0.14.0",
+    "langmem": "0.0.30",
+    "langgraph-checkpoint-postgres": "3.1.2",
+    "openai": "1.109.1",
+    "langchain-openai": "1.1.9"
+  },
+  "pinned_models": {
+    "chat": "gpt-4.1-mini-2025-04-14",
+    "embed": "text-embedding-3-small"
+  },
+  "ceilings_usd": {
+    "mem0-oss": 1.0,
+    "letta": 1.0,
+    "langmem": 1.0,
+    "zep": 0.0,
+    "total": 3.0
+  },
+  "modes": [
+    {
+      "mode": "smoke",
+      "task": "dup01_idempotent_remember",
+      "lanes": [
+        {
+          "lane": "letta",
+          "status": "completed",
+          "mode": "smoke",
+          "run_id": "opmem-phase2-letta-smoke",
+          "elapsed_seconds": 0.46,
+          "infrastructure_errors": 0,
+          "tasks_run": 1,
+          "tasks_passed": 1,
+          "task_results": [
+            {
+              "task": "dup01_idempotent_remember",
+              "passed": true,
+              "failures": []
+            }
+          ],
+          "product_constraints": []
+        }
+      ]
+    }
+  ]
+}
diff --git a/docs/benchmarks/artifacts/opmem-external-lanes/run-manifest-opmem-phase2-smoke1.json b/docs/benchmarks/artifacts/opmem-external-lanes/run-manifest-opmem-phase2-smoke1.json
new file mode 100644
index 00000000..381daa44
--- /dev/null
+++ b/docs/benchmarks/artifacts/opmem-external-lanes/run-manifest-opmem-phase2-smoke1.json
@@ -0,0 +1,86 @@
+{
+  "benchmark": "opmem-v0",
+  "generated_utc": "2026-09-19T08:28:45Z",
+  "run_id": "opmem-phase2-smoke1",
+  "host": "Linux-6.12.94+-x86_64-with-glibc2.39",
+  "python": "3.12.3",
+  "package_versions": {
+    "mem0ai": "2.1.0",
+    "qdrant-client": "1.19.1",
+    "zep-cloud": "3.28.0",
+    "tiktoken": "0.14.0",
+    "langmem": "0.0.30",
+    "langgraph-checkpoint-postgres": "3.1.2",
+    "openai": "3.16.2",
+    "langchain-openai": "1.6.2"
+  },
+  "pinned_models": {
+    "chat": "gpt-4.1-mini-2025-04-14",
+    "embed": "text-embedding-3-small"
+  },
+  "ceilings_usd": {
+    "mem0-oss": 1.0,
+    "letta": 1.0,
+    "langmem": 1.0,
+    "zep": 0.0,
+    "total": 3.0
+  },
+  "modes": [
+    {
+      "mode": "smoke",
+      "task": "dup01_idempotent_remember",
+      "lanes": [
+        {
+          "lane": "mem0-oss",
+          "status": "invalid_incomplete",
+          "mode": "smoke",
+          "run_id": "opmem-phase2-smoke1",
+          "elapsed_seconds": 6.36,
+          "infrastructure_errors": 1,
+          "tasks_run": 1,
+          "tasks_passed": 0,
+          "task_results": [
+            {
+              "task": "dup01_idempotent_remember",
+              "error": "ValueError: Top-level entity parameters frozenset({'user_id'}) are not supported in search(). Use filters={'user_id': '...'} instead.",
+              "invalid_incomplete": true
+            }
+          ],
+          "product_constraints": []
+        },
+        {
+          "lane": "langmem",
+          "status": "invalid_incomplete",
+          "mode": "smoke",
+          "run_id": "opmem-phase2-smoke1",
+          "elapsed_seconds": 3.15,
+          "infrastructure_errors": 1,
+          "tasks_run": 1,
+          "tasks_passed": 0,
+          "task_results": [
+            {
+              "task": "dup01_idempotent_remember",
+              "error": "AttributeError: 'dict' object has no attribute 'lower'",
+              "invalid_incomplete": true
+            }
+          ],
+          "product_constraints": []
+        },
+        {
+          "lane": "zep",
+          "status": "blocked",
+          "reason": "ZEP_API_KEY not present in runtime (vault not injected on this VM)",
+          "infrastructure_errors": 0,
+          "task_results": []
+        },
+        {
+          "lane": "letta",
+          "status": "blocked",
+          "reason": "LETTA_APP_SERVER_TOKEN not present in runtime",
+          "infrastructure_errors": 0,
+          "task_results": []
+        }
+      ]
+    }
+  ]
+}
diff --git a/docs/benchmarks/artifacts/opmem-external-lanes/run-manifest-opmem-phase2-smoke2.json b/docs/benchmarks/artifacts/opmem-external-lanes/run-manifest-opmem-phase2-smoke2.json
new file mode 100644
index 00000000..b712fde4
--- /dev/null
+++ b/docs/benchmarks/artifacts/opmem-external-lanes/run-manifest-opmem-phase2-smoke2.json
@@ -0,0 +1,72 @@
+{
+  "benchmark": "opmem-v0",
+  "generated_utc": "2026-09-19T08:29:11Z",
+  "run_id": "opmem-phase2-smoke2",
+  "host": "Linux-6.12.94+-x86_64-with-glibc2.39",
+  "python": "3.12.3",
+  "package_versions": {
+    "mem0ai": "2.1.0",
+    "qdrant-client": "1.19.1",
+    "zep-cloud": "3.28.0",
+    "tiktoken": "0.14.0",
+    "langmem": "0.0.30",
+    "langgraph-checkpoint-postgres": "3.1.2",
+    "openai": "3.16.2",
+    "langchain-openai": "1.6.2"
+  },
+  "pinned_models": {
+    "chat": "gpt-4.1-mini-2025-04-14",
+    "embed": "text-embedding-3-small"
+  },
+  "ceilings_usd": {
+    "mem0-oss": 1.0,
+    "letta": 1.0,
+    "langmem": 1.0,
+    "zep": 0.0,
+    "total": 3.0
+  },
+  "modes": [
+    {
+      "mode": "smoke",
+      "task": "dup01_idempotent_remember",
+      "lanes": [
+        {
+          "lane": "mem0-oss",
+          "status": "completed",
+          "mode": "smoke",
+          "run_id": "opmem-phase2-smoke2",
+          "elapsed_seconds": 2.9,
+          "infrastructure_errors": 0,
+          "tasks_run": 1,
+          "tasks_passed": 1,
+          "task_results": [
+            {
+              "task": "dup01_idempotent_remember",
+              "passed": true,
+              "failures": []
+            }
+          ],
+          "product_constraints": []
+        },
+        {
+          "lane": "langmem",
+          "status": "completed",
+          "mode": "smoke",
+          "run_id": "opmem-phase2-smoke2",
+          "elapsed_seconds": 3.68,
+          "infrastructure_errors": 0,
+          "tasks_run": 1,
+          "tasks_passed": 1,
+          "task_results": [
+            {
+              "task": "dup01_idempotent_remember",
+              "passed": true,
+              "failures": []
+            }
+          ],
+          "product_constraints": []
+        }
+      ]
+    }
+  ]
+}
diff --git a/docs/benchmarks/artifacts/opmem-external-lanes/run-manifest-opmem-phase2-zep-full.json b/docs/benchmarks/artifacts/opmem-external-lanes/run-manifest-opmem-phase2-zep-full.json
new file mode 100644
index 00000000..b7921084
--- /dev/null
+++ b/docs/benchmarks/artifacts/opmem-external-lanes/run-manifest-opmem-phase2-zep-full.json
@@ -0,0 +1,140 @@
+{
+  "benchmark": "opmem-v0",
+  "generated_utc": "2026-09-23T17:57:13Z",
+  "run_id": "opmem-phase2-zep-full",
+  "host": "Linux-6.12.94+-x86_64-with-glibc2.39",
+  "python": "3.12.3",
+  "package_versions": {
+    "mem0ai": "2.1.0",
+    "qdrant-client": "1.19.1",
+    "zep-cloud": "3.28.0",
+    "tiktoken": "0.14.0",
+    "langmem": "0.0.30",
+    "langgraph-checkpoint-postgres": "3.1.2",
+    "openai": "1.109.1",
+    "langchain-openai": "1.1.9"
+  },
+  "pinned_models": {
+    "chat": "gpt-4.1-mini-2025-04-14",
+    "embed": "text-embedding-3-small"
+  },
+  "ceilings_usd": {
+    "mem0-oss": 1.0,
+    "letta": 1.0,
+    "langmem": 1.0,
+    "zep": 0.0,
+    "total": 3.0
+  },
+  "modes": [
+    {
+      "mode": "full",
+      "task_count": 13,
+      "lanes": [
+        {
+          "lane": "zep",
+          "status": "completed",
+          "mode": "full",
+          "run_id": "opmem-phase2-zep-full",
+          "elapsed_seconds": 51.01,
+          "infrastructure_errors": 0,
+          "tasks_run": 13,
+          "tasks_passed": 3,
+          "task_results": [
+            {
+              "task": "cor01_basic_revision",
+              "passed": false,
+              "failures": [
+                "step1(recall): expected at least 1 results, got 0",
+                "step1(recall): top result does not contain 'python'"
+              ]
+            },
+            {
+              "task": "cor02_correction_stickiness",
+              "passed": true,
+              "failures": []
+            },
+            {
+              "task": "cor03_revised_retrievable",
+              "passed": true,
+              "failures": []
+            },
+            {
+              "task": "dup01_idempotent_remember",
+              "passed": true,
+              "failures": []
+            },
+            {
+              "task": "iso01_subject_isolation",
+              "passed": false,
+              "failures": [
+                "step1(recall): expected at most 0 results, got 1"
+              ]
+            },
+            {
+              "task": "iso02_tenant_isolation",
+              "passed": false,
+              "failures": [
+                "step1(recall): expected at most 0 results, got 1"
+              ]
+            },
+            {
+              "task": "iso03_forget_isolated",
+              "passed": false,
+              "failures": [
+                "step3(recall): expected at most 0 results, got 1"
+              ]
+            },
+            {
+              "task": "sup01_basic_forget",
+              "passed": false,
+              "failures": [
+                "step3(recall): expected at most 0 results, got 1"
+              ]
+            },
+            {
+              "task": "sup02_targeted_forget",
+              "passed": false,
+              "failures": [
+                "step3(recall): a result contains excluded 'door code'"
+              ]
+            },
+            {
+              "task": "sup03_durable_forget",
+              "passed": false,
+              "failures": [
+                "step3(recall): expected at most 0 results, got 2"
+              ]
+            },
+            {
+              "task": "upd01_stale_fact",
+              "passed": false,
+              "failures": [
+                "step2(recall): top result does not contain 'june'"
+              ]
+            },
+            {
+              "task": "upd02_preference_change",
+              "passed": false,
+              "failures": [
+                "step2(recall): expected at least 1 results, got 0",
+                "step2(recall): top result does not contain 'sms'"
+              ]
+            },
+            {
+              "task": "upd03_state_supersession",
+              "passed": false,
+              "failures": [
+                "step2(recall): top result does not contain '2222'"
+              ]
+            }
+          ],
+          "product_constraints": [
+            "lane zep cannot pin chat model gpt-4.1-mini-2025-04-14",
+            "lane zep cannot pin embed model text-embedding-3-small",
+            "Zep Cloud memory path does not accept OpenAI chat/embed model pins; stop before substituting \u2014 record scores only under Zep-native models."
+          ]
+        }
+      ]
+    }
+  ]
+}
diff --git a/docs/benchmarks/artifacts/opmem-external-lanes/run-manifest-opmem-phase2-zep-smoke-20260923.json b/docs/benchmarks/artifacts/opmem-external-lanes/run-manifest-opmem-phase2-zep-smoke-20260923.json
new file mode 100644
index 00000000..cfb9bcfc
--- /dev/null
+++ b/docs/benchmarks/artifacts/opmem-external-lanes/run-manifest-opmem-phase2-zep-smoke-20260923.json
@@ -0,0 +1,43 @@
+{
+  "benchmark": "opmem-v0",
+  "generated_utc": "2026-09-23T17:09:16Z",
+  "run_id": "opmem-phase2-zep-smoke-20260923",
+  "host": "Linux-6.12.94+-x86_64-with-glibc2.39",
+  "python": "3.12.3",
+  "package_versions": {
+    "mem0ai": "2.1.0",
+    "qdrant-client": "1.19.1",
+    "zep-cloud": "3.28.0",
+    "tiktoken": "0.14.0",
+    "langmem": "0.0.30",
+    "langgraph-checkpoint-postgres": "3.1.2",
+    "openai": "1.109.1",
+    "langchain-openai": "1.1.9"
+  },
+  "pinned_models": {
+    "chat": "gpt-4.1-mini-2025-04-14",
+    "embed": "text-embedding-3-small"
+  },
+  "ceilings_usd": {
+    "mem0-oss": 1.0,
+    "letta": 1.0,
+    "langmem": 1.0,
+    "zep": 0.0,
+    "total": 3.0
+  },
+  "modes": [
+    {
+      "mode": "smoke",
+      "task": "dup01_idempotent_remember",
+      "lanes": [
+        {
+          "lane": "zep",
+          "status": "blocked",
+          "reason": "ZEP_API_KEY not present in runtime (vault not injected on this VM)",
+          "infrastructure_errors": 0,
+          "task_results": []
+        }
+      ]
+    }
+  ]
+}
diff --git a/docs/benchmarks/artifacts/opmem-external-lanes/run-manifest-opmem-phase2-zep-smoke-20260923b.json b/docs/benchmarks/artifacts/opmem-external-lanes/run-manifest-opmem-phase2-zep-smoke-20260923b.json
new file mode 100644
index 00000000..a4c19363
--- /dev/null
+++ b/docs/benchmarks/artifacts/opmem-external-lanes/run-manifest-opmem-phase2-zep-smoke-20260923b.json
@@ -0,0 +1,58 @@
+{
+  "benchmark": "opmem-v0",
+  "generated_utc": "2026-09-23T17:55:34Z",
+  "run_id": "opmem-phase2-zep-smoke-20260923b",
+  "host": "Linux-6.12.94+-x86_64-with-glibc2.39",
+  "python": "3.12.3",
+  "package_versions": {
+    "mem0ai": "2.1.0",
+    "qdrant-client": "1.19.1",
+    "zep-cloud": "3.28.0",
+    "tiktoken": "0.14.0",
+    "langmem": "0.0.30",
+    "langgraph-checkpoint-postgres": "3.1.2",
+    "openai": "1.109.1",
+    "langchain-openai": "1.1.9"
+  },
+  "pinned_models": {
+    "chat": "gpt-4.1-mini-2025-04-14",
+    "embed": "text-embedding-3-small"
+  },
+  "ceilings_usd": {
+    "mem0-oss": 1.0,
+    "letta": 1.0,
+    "langmem": 1.0,
+    "zep": 0.0,
+    "total": 3.0
+  },
+  "modes": [
+    {
+      "mode": "smoke",
+      "task": "dup01_idempotent_remember",
+      "lanes": [
+        {
+          "lane": "zep",
+          "status": "invalid_incomplete",
+          "mode": "smoke",
+          "run_id": "opmem-phase2-zep-smoke-20260923b",
+          "elapsed_seconds": 0.32,
+          "infrastructure_errors": 1,
+          "tasks_run": 1,
+          "tasks_passed": 0,
+          "task_results": [
+            {
+              "task": "dup01_idempotent_remember",
+              "error": "AttributeError: 'Zep' object has no attribute 'memory'",
+              "invalid_incomplete": true
+            }
+          ],
+          "product_constraints": [
+            "lane zep cannot pin chat model gpt-4.1-mini-2025-04-14",
+            "lane zep cannot pin embed model text-embedding-3-small",
+            "Zep Cloud memory path does not accept OpenAI chat/embed model pins; stop before substituting \u2014 record scores only under Zep-native models."
+          ]
+        }
+      ]
+    }
+  ]
+}
diff --git a/docs/benchmarks/artifacts/opmem-external-lanes/run-manifest-opmem-phase2-zep-smoke-20260923c.json b/docs/benchmarks/artifacts/opmem-external-lanes/run-manifest-opmem-phase2-zep-smoke-20260923c.json
new file mode 100644
index 00000000..9967cebc
--- /dev/null
+++ b/docs/benchmarks/artifacts/opmem-external-lanes/run-manifest-opmem-phase2-zep-smoke-20260923c.json
@@ -0,0 +1,60 @@
+{
+  "benchmark": "opmem-v0",
+  "generated_utc": "2026-09-23T17:56:23Z",
+  "run_id": "opmem-phase2-zep-smoke-20260923c",
+  "host": "Linux-6.12.94+-x86_64-with-glibc2.39",
+  "python": "3.12.3",
+  "package_versions": {
+    "mem0ai": "2.1.0",
+    "qdrant-client": "1.19.1",
+    "zep-cloud": "3.28.0",
+    "tiktoken": "0.14.0",
+    "langmem": "0.0.30",
+    "langgraph-checkpoint-postgres": "3.1.2",
+    "openai": "1.109.1",
+    "langchain-openai": "1.1.9"
+  },
+  "pinned_models": {
+    "chat": "gpt-4.1-mini-2025-04-14",
+    "embed": "text-embedding-3-small"
+  },
+  "ceilings_usd": {
+    "mem0-oss": 1.0,
+    "letta": 1.0,
+    "langmem": 1.0,
+    "zep": 0.0,
+    "total": 3.0
+  },
+  "modes": [
+    {
+      "mode": "smoke",
+      "task": "dup01_idempotent_remember",
+      "lanes": [
+        {
+          "lane": "zep",
+          "status": "completed",
+          "mode": "smoke",
+          "run_id": "opmem-phase2-zep-smoke-20260923c",
+          "elapsed_seconds": 3.95,
+          "infrastructure_errors": 0,
+          "tasks_run": 1,
+          "tasks_passed": 0,
+          "task_results": [
+            {
+              "task": "dup01_idempotent_remember",
+              "passed": false,
+              "failures": [
+                "step2(recall): expected at least 1 results, got 0"
+              ]
+            }
+          ],
+          "product_constraints": [
+            "lane zep cannot pin chat model gpt-4.1-mini-2025-04-14",
+            "lane zep cannot pin embed model text-embedding-3-small",
+            "Zep Cloud memory path does not accept OpenAI chat/embed model pins; stop before substituting \u2014 record scores only under Zep-native models."
+          ]
+        }
+      ]
+    }
+  ]
+}
diff --git a/docs/benchmarks/artifacts/opmem-external-lanes/run-manifest-opmem-phase2-zep-smoke-final.json b/docs/benchmarks/artifacts/opmem-external-lanes/run-manifest-opmem-phase2-zep-smoke-final.json
new file mode 100644
index 00000000..2fb7143f
--- /dev/null
+++ b/docs/benchmarks/artifacts/opmem-external-lanes/run-manifest-opmem-phase2-zep-smoke-final.json
@@ -0,0 +1,60 @@
+{
+  "benchmark": "opmem-v0",
+  "generated_utc": "2026-09-23T17:56:50Z",
+  "run_id": "opmem-phase2-zep-smoke-final",
+  "host": "Linux-6.12.94+-x86_64-with-glibc2.39",
+  "python": "3.12.3",
+  "package_versions": {
+    "mem0ai": "2.1.0",
+    "qdrant-client": "1.19.1",
+    "zep-cloud": "3.28.0",
+    "tiktoken": "0.14.0",
+    "langmem": "0.0.30",
+    "langgraph-checkpoint-postgres": "3.1.2",
+    "openai": "1.109.1",
+    "langchain-openai": "1.1.9"
+  },
+  "pinned_models": {
+    "chat": "gpt-4.1-mini-2025-04-14",
+    "embed": "text-embedding-3-small"
+  },
+  "ceilings_usd": {
+    "mem0-oss": 1.0,
+    "letta": 1.0,
+    "langmem": 1.0,
+    "zep": 0.0,
+    "total": 3.0
+  },
+  "modes": [
+    {
+      "mode": "smoke",
+      "task": "dup01_idempotent_remember",
+      "lanes": [
+        {
+          "lane": "zep",
+          "status": "completed",
+          "mode": "smoke",
+          "run_id": "opmem-phase2-zep-smoke-final",
+          "elapsed_seconds": 4.01,
+          "infrastructure_errors": 0,
+          "tasks_run": 1,
+          "tasks_passed": 0,
+          "task_results": [
+            {
+              "task": "dup01_idempotent_remember",
+              "passed": false,
+              "failures": [
+                "step2(recall): expected at most 1 results, got 2"
+              ]
+            }
+          ],
+          "product_constraints": [
+            "lane zep cannot pin chat model gpt-4.1-mini-2025-04-14",
+            "lane zep cannot pin embed model text-embedding-3-small",
+            "Zep Cloud memory path does not accept OpenAI chat/embed model pins; stop before substituting \u2014 record scores only under Zep-native models."
+          ]
+        }
+      ]
+    }
+  ]
+}
diff --git a/docs/benchmarks/artifacts/opmem-external-lanes/run-manifest-opmem-phase2-zep-smoke.json b/docs/benchmarks/artifacts/opmem-external-lanes/run-manifest-opmem-phase2-zep-smoke.json
new file mode 100644
index 00000000..407bfb15
--- /dev/null
+++ b/docs/benchmarks/artifacts/opmem-external-lanes/run-manifest-opmem-phase2-zep-smoke.json
@@ -0,0 +1,43 @@
+{
+  "benchmark": "opmem-v0",
+  "generated_utc": "2026-09-21T17:43:56Z",
+  "run_id": "opmem-phase2-zep-smoke",
+  "host": "Linux-6.12.94+-x86_64-with-glibc2.39",
+  "python": "3.12.3",
+  "package_versions": {
+    "mem0ai": "2.1.0",
+    "qdrant-client": "1.19.1",
+    "zep-cloud": "3.28.0",
+    "tiktoken": "0.14.0",
+    "langmem": "0.0.30",
+    "langgraph-checkpoint-postgres": "3.1.2",
+    "openai": "1.109.1",
+    "langchain-openai": "1.1.9"
+  },
+  "pinned_models": {
+    "chat": "gpt-4.1-mini-2025-04-14",
+    "embed": "text-embedding-3-small"
+  },
+  "ceilings_usd": {
+    "mem0-oss": 1.0,
+    "letta": 1.0,
+    "langmem": 1.0,
+    "zep": 0.0,
+    "total": 3.0
+  },
+  "modes": [
+    {
+      "mode": "smoke",
+      "task": "dup01_idempotent_remember",
+      "lanes": [
+        {
+          "lane": "zep",
+          "status": "blocked",
+          "reason": "ZEP_API_KEY not present in runtime (vault not injected on this VM)",
+          "infrastructure_errors": 0,
+          "task_results": []
+        }
+      ]
+    }
+  ]
+}
diff --git a/docs/benchmarks/artifacts/opmem-external-lanes/telemetry-langmem-full.json b/docs/benchmarks/artifacts/opmem-external-lanes/telemetry-langmem-full.json
new file mode 100644
index 00000000..3035c9b9
--- /dev/null
+++ b/docs/benchmarks/artifacts/opmem-external-lanes/telemetry-langmem-full.json
@@ -0,0 +1,247 @@
+{
+  "lane": "langmem",
+  "usage": {
+    "chat_input_tokens": 0,
+    "chat_output_tokens": 0,
+    "embed_tokens": 0,
+    "estimated_usd": 0.0,
+    "http_requests": 47
+  },
+  "events": [
+    {
+      "op": "remember",
+      "status": 200,
+      "body_preview": "langmem manager.invoke"
+    },
+    {
+      "op": "recall",
+      "status": 200,
+      "body_preview": "[Item(namespace=['opmem', 'opmem-ad77ba8f7f247b76'], key='ff038350-4658-4617-92d9-4c5553430656', value={'kind': 'Memory', 'content': {'content': 'The user prefers Python for scripting.'}}, created_at='2026-09-19T08:29:52.169754+00:00', updated_at='2026-09-19T08:29:52.169754+00:00', score=0.33778279090288055)]"
+    },
+    {
+      "op": "revise",
+      "status": 200,
+      "body_preview": "langmem revise invoke"
+    },
+    {
+      "op": "recall",
+      "status": 200,
+      "body_preview": "[Item(namespace=['opmem', 'opmem-ad77ba8f7f247b76'], key='ff038350-4658-4617-92d9-4c5553430656', value={'kind': 'Memory', 'content': {'content': 'The user prefers Ruby for scripting.'}}, created_at='2026-09-19T08:29:52.169754+00:00', updated_at='2026-09-19T08:29:54.174845+00:00', score=0.33776719150121126)]"
+    },
+    {
+      "op": "remember",
+      "status": 200,
+      "body_preview": "langmem manager.invoke"
+    },
+    {
+      "op": "revise",
+      "status": 200,
+      "body_preview": "langmem revise invoke"
+    },
+    {
+      "op": "remember",
+      "status": 200,
+      "body_preview": "langmem manager.invoke"
+    },
+    {
+      "op": "recall",
+      "status": 200,
+      "body_preview": "[Item(namespace=['opmem', 'opmem-59710611587956f7'], key='20fc00d2-bdbe-498b-9cc2-94ae6d6926bc', value={'kind': 'Memory', 'content': {'content': 'User prefers Python for scripting.'}}, created_at='2026-09-19T08:29:57.897057+00:00', updated_at='2026-09-19T08:30:00.641575+00:00', score=0.33639699263808853), Item(namespace=['opmem', 'opmem-59710611587956f7'], key='8cb44cc6-27e1-488b-9394-d695a6bc0365', value={'kind': 'Memory', 'content': {'content': 'User prefers Python for scripting.'}}, created_a"
+    },
+    {
+      "op": "remember",
+      "status": 200,
+      "body_preview": "langmem manager.invoke"
+    },
+    {
+      "op": "revise",
+      "status": 200,
+      "body_preview": "langmem revise invoke"
+    },
+    {
+      "op": "recall",
+      "status": 200,
+      "body_preview": "[Item(namespace=['opmem', 'opmem-d89f907bfefacf0a'], key='c801ad89-40f4-4470-bfc4-9fa5cc81c716', value={'kind': 'Memory', 'content': {'content': 'The user prefers Ruby for scripting.'}}, created_at='2026-09-19T08:30:02.205766+00:00', updated_at='2026-09-19T08:30:04.766813+00:00', score=0.39614996459102103)]"
+    },
+    {
+      "op": "remember",
+      "status": 200,
+      "body_preview": "langmem manager.invoke"
+    },
+    {
+      "op": "remember",
+      "status": 200,
+      "body_preview": "langmem manager.invoke"
+    },
+    {
+      "op": "recall",
+      "status": 200,
+      "body_preview": "[Item(namespace=['opmem', 'opmem-231328502e0e8c77'], key='cf5aa7fc-14e2-4b88-9f5d-b7fc78fa702c', value={'kind': 'Memory', 'content': {'content': \"The user's wifi network name is FalconNet.\"}}, created_at='2026-09-19T08:30:06.559814+00:00', updated_at='2026-09-19T08:30:06.559814+00:00', score=0.4307357902741695)]"
+    },
+    {
+      "op": "remember",
+      "status": 200,
+      "body_preview": "langmem manager.invoke"
+    },
+    {
+      "op": "recall",
+      "status": 200,
+      "body_preview": "[]"
+    },
+    {
+      "op": "recall",
+      "status": 200,
+      "body_preview": "[Item(namespace=['opmem', 'opmem-f8551362a879e5a5'], key='f7dacf51-98ed-4656-baac-525040d049b3', value={'kind': 'Memory', 'content': {'content': \"The user's name is Alice Chen.\"}}, created_at='2026-09-19T08:30:09.585092+00:00', updated_at='2026-09-19T08:30:09.585092+00:00', score=0.4185555945234981)]"
+    },
+    {
+      "op": "remember",
+      "status": 200,
+      "body_preview": "langmem manager.invoke"
+    },
+    {
+      "op": "recall",
+      "status": 200,
+      "body_preview": "[]"
+    },
+    {
+      "op": "recall",
+      "status": 200,
+      "body_preview": "[Item(namespace=['opmem', 'opmem-0455ead06fd45698'], key='dd6965ac-061d-4b4c-bcd0-be701870e9e4', value={'kind': 'Memory', 'content': {'content': 'The WiFi network name is FalconNet.'}}, created_at='2026-09-19T08:30:11.539575+00:00', updated_at='2026-09-19T08:30:11.539575+00:00', score=0.4365041995152231)]"
+    },
+    {
+      "op": "remember",
+      "status": 200,
+      "body_preview": "langmem manager.invoke"
+    },
+    {
+      "op": "remember",
+      "status": 200,
+      "body_preview": "langmem manager.invoke"
+    },
+    {
+      "op": "forget",
+      "status": 200,
+      "body_preview": "deleted ['m-3432cec89e27']"
+    },
+    {
+      "op": "recall",
+      "status": 200,
+      "body_preview": "[Item(namespace=['opmem', 'opmem-001fc67db28bc8f6'], key='105cdf9d-ab27-471c-9f18-c86f1f0f6ebc', value={'kind': 'Memory', 'content': {'content': 'Never share the door code with vendors.'}}, created_at='2026-09-19T08:30:13.157541+00:00', updated_at='2026-09-19T08:30:13.157541+00:00', score=0.4455942668279429)]"
+    },
+    {
+      "op": "recall",
+      "status": 200,
+      "body_preview": "[Item(namespace=['opmem', 'opmem-6ba67fb6e2bcd4a0'], key='7d84ba1f-5876-4885-8513-a87a0fa94c24', value={'kind': 'Memory', 'content': {'content': 'Do not share the door code with vendors.'}}, created_at='2026-09-19T08:30:14.262160+00:00', updated_at='2026-09-19T08:30:14.262160+00:00', score=0.4431479037244387)]"
+    },
+    {
+      "op": "remember",
+      "status": 200,
+      "body_preview": "langmem manager.invoke"
+    },
+    {
+      "op": "recall",
+      "status": 200,
+      "body_preview": "[Item(namespace=['opmem', 'opmem-0769c5496ae38e0b'], key='62a0c516-bc1d-4f30-8557-67215486aee8', value={'kind': 'Memory', 'content': {'content': 'The user has a rule to never share the door code with vendors.'}}, created_at='2026-09-19T08:30:15.814395+00:00', updated_at='2026-09-19T08:30:15.814395+00:00', score=0.38258729230241584)]"
+    },
+    {
+      "op": "forget",
+      "status": 200,
+      "body_preview": "deleted ['m-10416bc6bd04']"
+    },
+    {
+      "op": "recall",
+      "status": 200,
+      "body_preview": "[Item(namespace=['opmem', 'opmem-0769c5496ae38e0b'], key='62a0c516-bc1d-4f30-8557-67215486aee8', value={'kind': 'Memory', 'content': {'content': 'The user has a rule to never share the door code with vendors.'}}, created_at='2026-09-19T08:30:15.814395+00:00', updated_at='2026-09-19T08:30:15.814395+00:00', score=0.38258729230241584)]"
+    },
+    {
+      "op": "remember",
+      "status": 200,
+      "body_preview": "langmem manager.invoke"
+    },
+    {
+      "op": "remember",
+      "status": 200,
+      "body_preview": "langmem manager.invoke"
+    },
+    {
+      "op": "forget",
+      "status": 200,
+      "body_preview": "deleted ['m-d98a3e5fcba9']"
+    },
+    {
+      "op": "recall",
+      "status": 200,
+      "body_preview": "[Item(namespace=['opmem', 'opmem-00a5581691894547'], key='2fe0745c-e090-4abb-8ee8-476f82624da3', value={'kind': 'Memory', 'content': {'content': 'Never share the door code with vendors.'}}, created_at='2026-09-19T08:30:17.777570+00:00', updated_at='2026-09-19T08:30:17.777570+00:00', score=0.44552312577829234), Item(namespace=['opmem', 'opmem-00a5581691894547'], key='169a612a-7f1b-43bb-919c-1f952bf3018b', value={'kind': 'Memory', 'content': {'content': 'The wifi network name is FalconNet.'}}, cre"
+    },
+    {
+      "op": "recall",
+      "status": 200,
+      "body_preview": "[Item(namespace=['opmem', 'opmem-00a5581691894547'], key='169a612a-7f1b-43bb-919c-1f952bf3018b', value={'kind': 'Memory', 'content': {'content': 'The wifi network name is FalconNet.'}}, created_at='2026-09-19T08:30:19.181133+00:00', updated_at='2026-09-19T08:30:19.181133+00:00', score=0.44823255645278093), Item(namespace=['opmem', 'opmem-00a5581691894547'], key='2fe0745c-e090-4abb-8ee8-476f82624da3', value={'kind': 'Memory', 'content': {'content': 'Never share the door code with vendors.'}}, cre"
+    },
+    {
+      "op": "remember",
+      "status": 200,
+      "body_preview": "langmem manager.invoke"
+    },
+    {
+      "op": "forget",
+      "status": 200,
+      "body_preview": "deleted ['m-357804338ee6']"
+    },
+    {
+      "op": "remember",
+      "status": 200,
+      "body_preview": "langmem manager.invoke"
+    },
+    {
+      "op": "recall",
+      "status": 200,
+      "body_preview": "[Item(namespace=['opmem', 'opmem-d011d90d55ad1853'], key='567d190a-d2f0-412a-8a2a-62d23bd5802c', value={'kind': 'Memory', 'content': {'content': 'User directive: Never share the door code with vendors.'}}, created_at='2026-09-19T08:30:20.884100+00:00', updated_at='2026-09-19T08:30:20.884100+00:00', score=0.42092394769662933)]"
+    },
+    {
+      "op": "remember",
+      "status": 200,
+      "body_preview": "langmem manager.invoke"
+    },
+    {
+      "op": "remember",
+      "status": 200,
+      "body_preview": "langmem manager.invoke"
+    },
+    {
+      "op": "recall",
+      "status": 200,
+      "body_preview": "[Item(namespace=['opmem', 'opmem-80742538490187d6'], key='575335a6-b0ea-423a-824e-29ff7fab88d4', value={'kind': 'Memory', 'content': {'content': 'The launch date is June 3.'}}, created_at='2026-09-19T08:30:23.681623+00:00', updated_at='2026-09-19T08:30:25.608805+00:00', score=0.4816371075183463)]"
+    },
+    {
+      "op": "remember",
+      "status": 200,
+      "body_preview": "langmem manager.invoke"
+    },
+    {
+      "op": "remember",
+      "status": 200,
+      "body_preview": "langmem manager.invoke"
+    },
+    {
+      "op": "recall",
+      "status": 200,
+      "body_preview": "[Item(namespace=['opmem', 'opmem-f9ff0a9d8b5da8ad'], key='29100b9a-a18e-4adf-afce-04fa698148bb', value={'kind': 'Memory', 'content': {'content': 'User preference: prefers SMS updates.'}}, created_at='2026-09-19T08:30:29.174995+00:00', updated_at='2026-09-19T08:30:29.174995+00:00', score=0.30517559730557475), Item(namespace=['opmem', 'opmem-f9ff0a9d8b5da8ad'], key='7f6b6804-28ee-49f5-ae45-15e137a5d47a', value={'kind': 'Memory', 'content': {'content': 'User preference: prefers SMS updates.'}}, cre"
+    },
+    {
+      "op": "remember",
+      "status": 200,
+      "body_preview": "langmem manager.invoke"
+    },
+    {
+      "op": "remember",
+      "status": 200,
+      "body_preview": "langmem manager.invoke"
+    },
+    {
+      "op": "recall",
+      "status": 200,
+      "body_preview": "[Item(namespace=['opmem', 'opmem-45283803a0a23fec'], key='4037cfed-accb-40d6-877e-e7075c207a71', value={'kind': 'Memory', 'content': {'content': 'The door code is 2222.'}}, created_at='2026-09-19T08:30:30.507762+00:00', updated_at='2026-09-19T08:30:32.350165+00:00', score=0.4994123415529361)]"
+    }
+  ]
+}
diff --git a/docs/benchmarks/artifacts/opmem-external-lanes/telemetry-langmem-smoke.json b/docs/benchmarks/artifacts/opmem-external-lanes/telemetry-langmem-smoke.json
new file mode 100644
index 00000000..d9ef7c1a
--- /dev/null
+++ b/docs/benchmarks/artifacts/opmem-external-lanes/telemetry-langmem-smoke.json
@@ -0,0 +1,27 @@
+{
+  "lane": "langmem",
+  "usage": {
+    "chat_input_tokens": 0,
+    "chat_output_tokens": 0,
+    "embed_tokens": 0,
+    "estimated_usd": 0.0,
+    "http_requests": 3
+  },
+  "events": [
+    {
+      "op": "remember",
+      "status": 200,
+      "body_preview": "langmem manager.invoke"
+    },
+    {
+      "op": "remember",
+      "status": 200,
+      "body_preview": "langmem manager.invoke"
+    },
+    {
+      "op": "recall",
+      "status": 200,
+      "body_preview": "[Item(namespace=['opmem', 'opmem-4dcb416cf36390f7'], key='e0f473c9-41e7-4cda-9bb3-6fcfd7ddc597', value={'kind': 'Memory', 'content': {'content': 'The wifi network name is FalconNet.'}}, created_at='2026-09-19T08:29:16.603964+00:00', updated_at='2026-09-19T08:29:16.603964+00:00', score=0.44823255645278093)]"
+    }
+  ]
+}
diff --git a/docs/benchmarks/artifacts/opmem-external-lanes/telemetry-letta-full.json b/docs/benchmarks/artifacts/opmem-external-lanes/telemetry-letta-full.json
new file mode 100644
index 00000000..683701bd
--- /dev/null
+++ b/docs/benchmarks/artifacts/opmem-external-lanes/telemetry-letta-full.json
@@ -0,0 +1,347 @@
+{
+  "lane": "letta",
+  "usage": {
+    "chat_input_tokens": 0,
+    "chat_output_tokens": 0,
+    "embed_tokens": 0,
+    "estimated_usd": 0.0,
+    "http_requests": 67
+  },
+  "events": [
+    {
+      "op": "reset",
+      "status": 200,
+      "body_preview": "agent agent-0e765134-eef7-40be-bca4-fd55de9054e6"
+    },
+    {
+      "op": "blocks.modify",
+      "status": 200,
+      "body_preview": "1 facts"
+    },
+    {
+      "op": "recall",
+      "status": 200,
+      "body_preview": "[]"
+    },
+    {
+      "op": "blocks.modify",
+      "status": 200,
+      "body_preview": "1 facts"
+    },
+    {
+      "op": "revise",
+      "status": 200,
+      "body_preview": "updated m-72d0cb92d778"
+    },
+    {
+      "op": "recall",
+      "status": 200,
+      "body_preview": "[{'id': 'm-72d0cb92d778', 'content': 'Prefers Ruby for scripting'}]"
+    },
+    {
+      "op": "reset",
+      "status": 200,
+      "body_preview": "agent agent-1980a99b-fa48-4eea-bde4-fb8d923793cc"
+    },
+    {
+      "op": "blocks.modify",
+      "status": 200,
+      "body_preview": "1 facts"
+    },
+    {
+      "op": "blocks.modify",
+      "status": 200,
+      "body_preview": "1 facts"
+    },
+    {
+      "op": "revise",
+      "status": 200,
+      "body_preview": "updated m-9bdc773d176d"
+    },
+    {
+      "op": "blocks.modify",
+      "status": 200,
+      "body_preview": "1 facts"
+    },
+    {
+      "op": "recall",
+      "status": 200,
+      "body_preview": "[]"
+    },
+    {
+      "op": "reset",
+      "status": 200,
+      "body_preview": "agent agent-e32f945a-47de-4061-afff-882d2c175d82"
+    },
+    {
+      "op": "blocks.modify",
+      "status": 200,
+      "body_preview": "1 facts"
+    },
+    {
+      "op": "blocks.modify",
+      "status": 200,
+      "body_preview": "1 facts"
+    },
+    {
+      "op": "revise",
+      "status": 200,
+      "body_preview": "updated m-e53c02783ebf"
+    },
+    {
+      "op": "recall",
+      "status": 200,
+      "body_preview": "[{'id': 'm-e53c02783ebf', 'content': 'Prefers Ruby for scripting'}]"
+    },
+    {
+      "op": "reset",
+      "status": 200,
+      "body_preview": "agent agent-d6215054-0208-4946-a9cf-9e3263da81c2"
+    },
+    {
+      "op": "blocks.modify",
+      "status": 200,
+      "body_preview": "1 facts"
+    },
+    {
+      "op": "blocks.modify",
+      "status": 200,
+      "body_preview": "1 facts"
+    },
+    {
+      "op": "recall",
+      "status": 200,
+      "body_preview": "[{'id': 'm-86f89e5fc162', 'content': 'The wifi network name is FalconNet.'}]"
+    },
+    {
+      "op": "reset",
+      "status": 200,
+      "body_preview": "agent agent-e15bf9de-e27a-4c65-815d-4da1f2b49de0"
+    },
+    {
+      "op": "blocks.modify",
+      "status": 200,
+      "body_preview": "1 facts"
+    },
+    {
+      "op": "recall",
+      "status": 200,
+      "body_preview": "[{'id': 'm-3d6f2ccfc041', 'content': 'My name is Alice Chen.'}]"
+    },
+    {
+      "op": "recall",
+      "status": 200,
+      "body_preview": "[{'id': 'm-3d6f2ccfc041', 'content': 'My name is Alice Chen.'}]"
+    },
+    {
+      "op": "reset",
+      "status": 200,
+      "body_preview": "agent agent-dc86f6bd-b17c-4219-badd-1e19389c0c91"
+    },
+    {
+      "op": "blocks.modify",
+      "status": 200,
+      "body_preview": "1 facts"
+    },
+    {
+      "op": "recall",
+      "status": 200,
+      "body_preview": "[{'id': 'm-6d39288e1285', 'content': 'The wifi network name is FalconNet.'}]"
+    },
+    {
+      "op": "recall",
+      "status": 200,
+      "body_preview": "[{'id': 'm-6d39288e1285', 'content': 'The wifi network name is FalconNet.'}]"
+    },
+    {
+      "op": "reset",
+      "status": 200,
+      "body_preview": "agent agent-2c51ef1a-b9b5-407c-9c25-ded27519a2a0"
+    },
+    {
+      "op": "blocks.modify",
+      "status": 200,
+      "body_preview": "1 facts"
+    },
+    {
+      "op": "blocks.modify",
+      "status": 200,
+      "body_preview": "1 facts"
+    },
+    {
+      "op": "blocks.modify",
+      "status": 200,
+      "body_preview": "0 facts"
+    },
+    {
+      "op": "forget",
+      "status": 200,
+      "body_preview": "removed ['m-199792273a68']"
+    },
+    {
+      "op": "recall",
+      "status": 200,
+      "body_preview": "[]"
+    },
+    {
+      "op": "recall",
+      "status": 200,
+      "body_preview": "[]"
+    },
+    {
+      "op": "reset",
+      "status": 200,
+      "body_preview": "agent agent-8d05177a-19d2-4dbe-ad34-697cc6418823"
+    },
+    {
+      "op": "blocks.modify",
+      "status": 200,
+      "body_preview": "1 facts"
+    },
+    {
+      "op": "recall",
+      "status": 200,
+      "body_preview": "[{'id': 'm-fcff7a76adb5', 'content': 'Never share the door code with vendors.'}]"
+    },
+    {
+      "op": "blocks.modify",
+      "status": 200,
+      "body_preview": "0 facts"
+    },
+    {
+      "op": "forget",
+      "status": 200,
+      "body_preview": "removed ['m-fcff7a76adb5']"
+    },
+    {
+      "op": "recall",
+      "status": 200,
+      "body_preview": "[]"
+    },
+    {
+      "op": "reset",
+      "status": 200,
+      "body_preview": "agent agent-cd220203-8ef1-42e0-b207-b0c7076560b1"
+    },
+    {
+      "op": "blocks.modify",
+      "status": 200,
+      "body_preview": "1 facts"
+    },
+    {
+      "op": "blocks.modify",
+      "status": 200,
+      "body_preview": "2 facts"
+    },
+    {
+      "op": "blocks.modify",
+      "status": 200,
+      "body_preview": "1 facts"
+    },
+    {
+      "op": "forget",
+      "status": 200,
+      "body_preview": "removed ['m-876d3c78fdbb']"
+    },
+    {
+      "op": "recall",
+      "status": 200,
+      "body_preview": "[]"
+    },
+    {
+      "op": "recall",
+      "status": 200,
+      "body_preview": "[{'id': 'm-8a2ae350d5a4', 'content': 'The wifi network name is FalconNet.'}]"
+    },
+    {
+      "op": "reset",
+      "status": 200,
+      "body_preview": "agent agent-91dd1cc4-65a7-4b6f-a5e7-fff67b8612f8"
+    },
+    {
+      "op": "blocks.modify",
+      "status": 200,
+      "body_preview": "1 facts"
+    },
+    {
+      "op": "blocks.modify",
+      "status": 200,
+      "body_preview": "0 facts"
+    },
+    {
+      "op": "forget",
+      "status": 200,
+      "body_preview": "removed ['m-f372883003fe']"
+    },
+    {
+      "op": "blocks.modify",
+      "status": 200,
+      "body_preview": "1 facts"
+    },
+    {
+      "op": "recall",
+      "status": 200,
+      "body_preview": "[{'id': 'm-f372883003fe', 'content': 'Never share the door code with vendors.'}]"
+    },
+    {
+      "op": "reset",
+      "status": 200,
+      "body_preview": "agent agent-84ee20a4-8121-4b38-aa77-d6f23aaaa6b6"
+    },
+    {
+      "op": "blocks.modify",
+      "status": 200,
+      "body_preview": "1 facts"
+    },
+    {
+      "op": "blocks.modify",
+      "status": 200,
+      "body_preview": "2 facts"
+    },
+    {
+      "op": "recall",
+      "status": 200,
+      "body_preview": "[{'id': 'm-539fd5974dad', 'content': 'The launch date is May 12.'}, {'id': 'm-a3552df98ca5', 'content': 'The launch date is June 3.'}]"
+    },
+    {
+      "op": "reset",
+      "status": 200,
+      "body_preview": "agent agent-bcbd0fe3-f78d-497e-a2fd-c79d2331e434"
+    },
+    {
+      "op": "blocks.modify",
+      "status": 200,
+      "body_preview": "1 facts"
+    },
+    {
+      "op": "blocks.modify",
+      "status": 200,
+      "body_preview": "2 facts"
+    },
+    {
+      "op": "recall",
+      "status": 200,
+      "body_preview": "[]"
+    },
+    {
+      "op": "reset",
+      "status": 200,
+      "body_preview": "agent agent-019571c6-1f7d-40e1-9300-8e7d32de02a2"
+    },
+    {
+      "op": "blocks.modify",
+      "status": 200,
+      "body_preview": "1 facts"
+    },
+    {
+      "op": "blocks.modify",
+      "status": 200,
+      "body_preview": "2 facts"
+    },
+    {
+      "op": "recall",
+      "status": 200,
+      "body_preview": "[{'id': 'm-6d0ba2632157', 'content': 'The door code is 2222.'}, {'id': 'm-f216d185ca5e', 'content': 'The door code is 1111.'}]"
+    }
+  ]
+}
diff --git a/docs/benchmarks/artifacts/opmem-external-lanes/telemetry-letta-smoke.json b/docs/benchmarks/artifacts/opmem-external-lanes/telemetry-letta-smoke.json
new file mode 100644
index 00000000..8d051e29
--- /dev/null
+++ b/docs/benchmarks/artifacts/opmem-external-lanes/telemetry-letta-smoke.json
@@ -0,0 +1,32 @@
+{
+  "lane": "letta",
+  "usage": {
+    "chat_input_tokens": 0,
+    "chat_output_tokens": 0,
+    "embed_tokens": 0,
+    "estimated_usd": 0.0,
+    "http_requests": 4
+  },
+  "events": [
+    {
+      "op": "reset",
+      "status": 200,
+      "body_preview": "agent agent-eefd561c-670e-46e2-a634-17a5cb704ba2"
+    },
+    {
+      "op": "blocks.modify",
+      "status": 200,
+      "body_preview": "1 facts"
+    },
+    {
+      "op": "blocks.modify",
+      "status": 200,
+      "body_preview": "1 facts"
+    },
+    {
+      "op": "recall",
+      "status": 200,
+      "body_preview": "[{'id': 'm-5ce49f74b219', 'content': 'The wifi network name is FalconNet.'}]"
+    }
+  ]
+}
diff --git a/docs/benchmarks/artifacts/opmem-external-lanes/telemetry-mem0-oss-full.json b/docs/benchmarks/artifacts/opmem-external-lanes/telemetry-mem0-oss-full.json
new file mode 100644
index 00000000..8ef89380
--- /dev/null
+++ b/docs/benchmarks/artifacts/opmem-external-lanes/telemetry-mem0-oss-full.json
@@ -0,0 +1,247 @@
+{
+  "lane": "mem0-oss",
+  "usage": {
+    "chat_input_tokens": 0,
+    "chat_output_tokens": 0,
+    "embed_tokens": 0,
+    "estimated_usd": 0.0,
+    "http_requests": 47
+  },
+  "events": [
+    {
+      "op": "remember",
+      "status": 200,
+      "body_preview": "{'results': [{'id': '5c8819cd-ce54-413d-9e88-03ecd4f70f52', 'memory': 'User prefers Python for scripting tasks', 'event': 'ADD'}]}"
+    },
+    {
+      "op": "recall",
+      "status": 200,
+      "body_preview": "{'results': [{'id': '5c8819cd-ce54-413d-9e88-03ecd4f70f52', 'memory': 'User prefers Python for scripting tasks', 'hash': 'a48dc9b92de4e17c1cc8962e11c6dd48', 'metadata': None, 'score': 0.47747782832465313, 'created_at': '2026-09-19T08:30:48.595042+00:00', 'updated_at': '2026-09-19T08:30:48.595042+00:00', 'user_id': 'opmem-5e7cd8f9c26a01c7', 'attributed_to': 'user'}]}"
+    },
+    {
+      "op": "revise",
+      "status": 200,
+      "body_preview": "updated 5c8819cd-ce54-413d-9e88-03ecd4f70f52"
+    },
+    {
+      "op": "recall",
+      "status": 200,
+      "body_preview": "{'results': [{'id': '5c8819cd-ce54-413d-9e88-03ecd4f70f52', 'memory': 'Prefers Ruby for scripting', 'hash': 'f6e4b4415f448d039530b6671977e0b2', 'metadata': None, 'score': 0.5080760678115537, 'created_at': '2026-09-19T08:30:48.595042+00:00', 'updated_at': '2026-09-19T08:30:48.935555+00:00', 'user_id': 'opmem-5e7cd8f9c26a01c7', 'attributed_to': 'user'}]}"
+    },
+    {
+      "op": "remember",
+      "status": 200,
+      "body_preview": "{'results': [{'id': '7a05a16c-92e8-4b02-bfe7-c0aaa4fb865b', 'memory': 'User prefers Python for scripting tasks', 'event': 'ADD'}]}"
+    },
+    {
+      "op": "revise",
+      "status": 200,
+      "body_preview": "updated 7a05a16c-92e8-4b02-bfe7-c0aaa4fb865b"
+    },
+    {
+      "op": "remember",
+      "status": 200,
+      "body_preview": "{'results': [{'id': '2e2eee9d-57d6-430d-8aa8-7255274580f1', 'memory': 'User prefers Python for scripting, updating their previous preference for Ruby', 'event': 'ADD'}]}"
+    },
+    {
+      "op": "recall",
+      "status": 200,
+      "body_preview": "{'results': [{'id': '7a05a16c-92e8-4b02-bfe7-c0aaa4fb865b', 'memory': 'Prefers Ruby for scripting', 'hash': 'f6e4b4415f448d039530b6671977e0b2', 'metadata': None, 'score': 0.5080626159016706, 'created_at': '2026-09-19T08:30:52.277528+00:00', 'updated_at': '2026-09-19T08:30:52.458560+00:00', 'user_id': 'opmem-191998c3fb442640', 'attributed_to': 'user'}, {'id': '2e2eee9d-57d6-430d-8aa8-7255274580f1', 'memory': 'User prefers Python for scripting, updating their previous preference for Ruby', 'hash':"
+    },
+    {
+      "op": "remember",
+      "status": 200,
+      "body_preview": "{'results': [{'id': 'be262d77-4d24-4228-b910-1bf10000435e', 'memory': 'User prefers Python for scripting tasks.', 'event': 'ADD'}]}"
+    },
+    {
+      "op": "revise",
+      "status": 200,
+      "body_preview": "updated be262d77-4d24-4228-b910-1bf10000435e"
+    },
+    {
+      "op": "recall",
+      "status": 200,
+      "body_preview": "{'results': [{'id': 'be262d77-4d24-4228-b910-1bf10000435e', 'memory': 'Prefers Ruby for scripting', 'hash': 'f6e4b4415f448d039530b6671977e0b2', 'metadata': None, 'score': 0.5780423781379709, 'created_at': '2026-09-19T08:30:57.173437+00:00', 'updated_at': '2026-09-19T08:30:57.374013+00:00', 'user_id': 'opmem-59047463df9dc059', 'attributed_to': 'user'}]}"
+    },
+    {
+      "op": "remember",
+      "status": 200,
+      "body_preview": "{'results': [{'id': '9388e553-6720-4596-828a-1364e2125f4c', 'memory': \"User's wifi network name is FalconNet\", 'event': 'ADD'}]}"
+    },
+    {
+      "op": "remember",
+      "status": 200,
+      "body_preview": "{'results': []}"
+    },
+    {
+      "op": "recall",
+      "status": 200,
+      "body_preview": "{'results': [{'id': '9388e553-6720-4596-828a-1364e2125f4c', 'memory': \"User's wifi network name is FalconNet\", 'hash': '0e881d9cc295e8509eb50d959ed86653', 'metadata': None, 'score': 0.587925614386172, 'created_at': '2026-09-19T08:31:00.593416+00:00', 'updated_at': '2026-09-19T08:31:00.593416+00:00', 'user_id': 'opmem-791c09301b2c6c03', 'attributed_to': 'user'}]}"
+    },
+    {
+      "op": "remember",
+      "status": 200,
+      "body_preview": "{'results': [{'id': '5554baa6-3ca9-47b9-8515-fe42cf7d1a95', 'memory': \"User's name is Alice Chen\", 'event': 'ADD'}]}"
+    },
+    {
+      "op": "recall",
+      "status": 200,
+      "body_preview": "{'results': []}"
+    },
+    {
+      "op": "recall",
+      "status": 200,
+      "body_preview": "{'results': [{'id': '5554baa6-3ca9-47b9-8515-fe42cf7d1a95', 'memory': \"User's name is Alice Chen\", 'hash': 'b7c543b1a5c7edd3d2a5bed7ad46fb84', 'metadata': None, 'score': 0.6258135863190433, 'created_at': '2026-09-19T08:31:02.655875+00:00', 'updated_at': '2026-09-19T08:31:02.655875+00:00', 'user_id': 'opmem-c659fde1c729f26a', 'attributed_to': 'user'}]}"
+    },
+    {
+      "op": "remember",
+      "status": 200,
+      "body_preview": "{'results': [{'id': 'fc780410-3627-4a3f-9b7e-3083586df000', 'memory': \"User's wifi network name is FalconNet\", 'event': 'ADD'}]}"
+    },
+    {
+      "op": "recall",
+      "status": 200,
+      "body_preview": "{'results': []}"
+    },
+    {
+      "op": "recall",
+      "status": 200,
+      "body_preview": "{'results': [{'id': 'fc780410-3627-4a3f-9b7e-3083586df000', 'memory': \"User's wifi network name is FalconNet\", 'hash': '0e881d9cc295e8509eb50d959ed86653', 'metadata': None, 'score': 0.587925614386172, 'created_at': '2026-09-19T08:31:04.048007+00:00', 'updated_at': '2026-09-19T08:31:04.048007+00:00', 'user_id': 'opmem-fc2994ec968c8c35', 'attributed_to': 'user'}]}"
+    },
+    {
+      "op": "remember",
+      "status": 200,
+      "body_preview": "{'results': [{'id': '8a4933cc-883e-4524-a2c1-6f2bb9e2b0cd', 'memory': 'User instructs never to share the door code with vendors to maintain security.', 'event': 'ADD'}]}"
+    },
+    {
+      "op": "remember",
+      "status": 200,
+      "body_preview": "{'results': [{'id': '77a39bcc-4cd3-44b1-9430-49f57c87496d', 'memory': 'User instructs never to share the door code with vendors to maintain security', 'event': 'ADD'}]}"
+    },
+    {
+      "op": "forget",
+      "status": 200,
+      "body_preview": "deleted 8a4933cc-883e-4524-a2c1-6f2bb9e2b0cd"
+    },
+    {
+      "op": "recall",
+      "status": 200,
+      "body_preview": "{'results': []}"
+    },
+    {
+      "op": "recall",
+      "status": 200,
+      "body_preview": "{'results': [{'id': '77a39bcc-4cd3-44b1-9430-49f57c87496d', 'memory': 'User instructs never to share the door code with vendors to maintain security', 'hash': '5eef4b9cdafb75fc82e5467f75e2dd30', 'metadata': None, 'score': 0.555903332358419, 'created_at': '2026-09-19T08:31:06.847614+00:00', 'updated_at': '2026-09-19T08:31:06.847614+00:00', 'user_id': 'opmem-3c9b018485c2a47c', 'attributed_to': 'user'}]}"
+    },
+    {
+      "op": "remember",
+      "status": 200,
+      "body_preview": "{'results': [{'id': '5bb8d9b0-3d68-4c2e-8014-5613f99bbc39', 'memory': 'User has a strict rule to never share the door code with vendors to maintain security.', 'event': 'ADD'}]}"
+    },
+    {
+      "op": "recall",
+      "status": 200,
+      "body_preview": "{'results': [{'id': '5bb8d9b0-3d68-4c2e-8014-5613f99bbc39', 'memory': 'User has a strict rule to never share the door code with vendors to maintain security.', 'hash': 'd2a86a6f638f37284a6e9e69bc78b9b8', 'metadata': None, 'score': 0.5078514150420694, 'created_at': '2026-09-19T08:31:08.562864+00:00', 'updated_at': '2026-09-19T08:31:08.562864+00:00', 'user_id': 'opmem-e915b53e598654e2', 'attributed_to': 'user'}]}"
+    },
+    {
+      "op": "forget",
+      "status": 200,
+      "body_preview": "deleted 5bb8d9b0-3d68-4c2e-8014-5613f99bbc39"
+    },
+    {
+      "op": "recall",
+      "status": 200,
+      "body_preview": "{'results': []}"
+    },
+    {
+      "op": "remember",
+      "status": 200,
+      "body_preview": "{'results': [{'id': '74a731ad-891b-44a5-8d47-524a7a69bcc3', 'memory': 'User instructs never to share the door code with vendors to maintain security', 'event': 'ADD'}]}"
+    },
+    {
+      "op": "remember",
+      "status": 200,
+      "body_preview": "{'results': [{'id': 'abbd8373-beca-4604-a351-c869573bed28', 'memory': \"User's wifi network name is FalconNet\", 'event': 'ADD'}]}"
+    },
+    {
+      "op": "forget",
+      "status": 200,
+      "body_preview": "deleted 74a731ad-891b-44a5-8d47-524a7a69bcc3"
+    },
+    {
+      "op": "recall",
+      "status": 200,
+      "body_preview": "{'results': [{'id': 'abbd8373-beca-4604-a351-c869573bed28', 'memory': \"User's wifi network name is FalconNet\", 'hash': '0e881d9cc295e8509eb50d959ed86653', 'metadata': None, 'score': 0.2068300875562672, 'created_at': '2026-09-19T08:31:11.421207+00:00', 'updated_at': '2026-09-19T08:31:11.421207+00:00', 'user_id': 'opmem-327e9acda69047a2', 'attributed_to': 'user'}]}"
+    },
+    {
+      "op": "recall",
+      "status": 200,
+      "body_preview": "{'results': [{'id': 'abbd8373-beca-4604-a351-c869573bed28', 'memory': \"User's wifi network name is FalconNet\", 'hash': '0e881d9cc295e8509eb50d959ed86653', 'metadata': None, 'score': 0.5879256143861721, 'created_at': '2026-09-19T08:31:11.421207+00:00', 'updated_at': '2026-09-19T08:31:11.421207+00:00', 'user_id': 'opmem-327e9acda69047a2', 'attributed_to': 'user'}]}"
+    },
+    {
+      "op": "remember",
+      "status": 200,
+      "body_preview": "{'results': [{'id': 'd6481591-450e-4a48-8b08-826548c9cd88', 'memory': 'User instructs never to share the door code with vendors to maintain security.', 'event': 'ADD'}]}"
+    },
+    {
+      "op": "forget",
+      "status": 200,
+      "body_preview": "deleted d6481591-450e-4a48-8b08-826548c9cd88"
+    },
+    {
+      "op": "remember",
+      "status": 200,
+      "body_preview": "{'results': [{'id': 'ef10248a-f45e-4f2e-9e0b-2004db58a01a', 'memory': 'User has a strict rule to never share the door code with vendors for security reasons', 'event': 'ADD'}]}"
+    },
+    {
+      "op": "recall",
+      "status": 200,
+      "body_preview": "{'results': [{'id': 'ef10248a-f45e-4f2e-9e0b-2004db58a01a', 'memory': 'User has a strict rule to never share the door code with vendors for security reasons', 'hash': 'fbcafb4b09ccf7e171a380515f78367a', 'metadata': None, 'score': 0.5323019705553599, 'created_at': '2026-09-19T08:31:14.079406+00:00', 'updated_at': '2026-09-19T08:31:14.079406+00:00', 'user_id': 'opmem-e929007e8f1422c5', 'attributed_to': 'user'}]}"
+    },
+    {
+      "op": "remember",
+      "status": 200,
+      "body_preview": "{'results': [{'id': '355365f8-ebef-4d4d-b914-612e7bf7083e', 'memory': 'User stated that the launch date is May 12', 'event': 'ADD'}]}"
+    },
+    {
+      "op": "remember",
+      "status": 200,
+      "body_preview": "{'results': [{'id': '97cd9fb1-8a12-465b-8160-0a47bc7d4667', 'memory': 'User updated the launch date to June 3, 2026, replacing the previous date of May 12, 2026', 'event': 'ADD'}]}"
+    },
+    {
+      "op": "recall",
+      "status": 200,
+      "body_preview": "{'results': [{'id': '355365f8-ebef-4d4d-b914-612e7bf7083e', 'memory': 'User stated that the launch date is May 12', 'hash': 'fa737facb798b3f3fc1535dc256ebf00', 'metadata': None, 'score': 0.6070983935080743, 'created_at': '2026-09-19T08:31:15.501751+00:00', 'updated_at': '2026-09-19T08:31:15.501751+00:00', 'user_id': 'opmem-e98de152102dc08d', 'attributed_to': 'user'}, {'id': '97cd9fb1-8a12-465b-8160-0a47bc7d4667', 'memory': 'User updated the launch date to June 3, 2026, replacing the previous dat"
+    },
+    {
+      "op": "remember",
+      "status": 200,
+      "body_preview": "{'results': [{'id': '5b51b507-b33e-408a-8a00-0958ef839ad2', 'memory': 'User prefers to receive updates via email', 'event': 'ADD'}]}"
+    },
+    {
+      "op": "remember",
+      "status": 200,
+      "body_preview": "{'results': [{'id': '73a3c476-f478-4c49-8163-1296e6546265', 'memory': 'User prefers SMS updates instead of email updates as of September 19, 2026', 'event': 'ADD'}]}"
+    },
+    {
+      "op": "recall",
+      "status": 200,
+      "body_preview": "{'results': [{'id': '5b51b507-b33e-408a-8a00-0958ef839ad2', 'memory': 'User prefers to receive updates via email', 'hash': '6bfa2475c3bc61a23a48fdff754bfd8d', 'metadata': None, 'score': 0.38536272449852116, 'created_at': '2026-09-19T08:31:18.395105+00:00', 'updated_at': '2026-09-19T08:31:18.395105+00:00', 'user_id': 'opmem-7a5832b838c04f48', 'attributed_to': 'user'}, {'id': '73a3c476-f478-4c49-8163-1296e6546265', 'memory': 'User prefers SMS updates instead of email updates as of September 19, 20"
+    },
+    {
+      "op": "remember",
+      "status": 200,
+      "body_preview": "{'results': [{'id': '34148019-1faf-4ea6-9762-704993a93489', 'memory': 'User shared that the door code is 1111', 'event': 'ADD'}]}"
+    },
+    {
+      "op": "remember",
+      "status": 200,
+      "body_preview": "{'results': [{'id': '77525f51-2be9-4376-8c70-3b12b05392cd', 'memory': 'User updated the door code to 2222 on September 19, 2026', 'event': 'ADD'}]}"
+    },
+    {
+      "op": "recall",
+      "status": 200,
+      "body_preview": "{'results': [{'id': '34148019-1faf-4ea6-9762-704993a93489', 'memory': 'User shared that the door code is 1111', 'hash': 'd4ff66858269b4e5157f79c4164a9ed1', 'metadata': None, 'score': 0.6118697230882042, 'created_at': '2026-09-19T08:31:20.904109+00:00', 'updated_at': '2026-09-19T08:31:20.904109+00:00', 'user_id': 'opmem-d3d362b2f352c734', 'attributed_to': 'user'}, {'id': '77525f51-2be9-4376-8c70-3b12b05392cd', 'memory': 'User updated the door code to 2222 on September 19, 2026', 'hash': '1ea2b7f3"
+    }
+  ]
+}
diff --git a/docs/benchmarks/artifacts/opmem-external-lanes/telemetry-mem0-oss-smoke.json b/docs/benchmarks/artifacts/opmem-external-lanes/telemetry-mem0-oss-smoke.json
new file mode 100644
index 00000000..edcb6dd6
--- /dev/null
+++ b/docs/benchmarks/artifacts/opmem-external-lanes/telemetry-mem0-oss-smoke.json
@@ -0,0 +1,27 @@
+{
+  "lane": "mem0-oss",
+  "usage": {
+    "chat_input_tokens": 0,
+    "chat_output_tokens": 0,
+    "embed_tokens": 0,
+    "estimated_usd": 0.0,
+    "http_requests": 3
+  },
+  "events": [
+    {
+      "op": "remember",
+      "status": 200,
+      "body_preview": "{'results': [{'id': '7d6f7716-8895-4a86-839f-25514b82599e', 'memory': \"User's wifi network name is FalconNet\", 'event': 'ADD'}]}"
+    },
+    {
+      "op": "remember",
+      "status": 200,
+      "body_preview": "{'results': []}"
+    },
+    {
+      "op": "recall",
+      "status": 200,
+      "body_preview": "{'results': [{'id': '7d6f7716-8895-4a86-839f-25514b82599e', 'memory': \"User's wifi network name is FalconNet\", 'hash': '0e881d9cc295e8509eb50d959ed86653', 'metadata': None, 'score': 0.587925614386172, 'created_at': '2026-09-19T08:29:14.005476+00:00', 'updated_at': '2026-09-19T08:29:14.005476+00:00', 'user_id': 'opmem-8d09a77b2f4c4647', 'attributed_to': 'user'}]}"
+    }
+  ]
+}
diff --git a/docs/benchmarks/artifacts/opmem-external-lanes/telemetry-zep-full.json b/docs/benchmarks/artifacts/opmem-external-lanes/telemetry-zep-full.json
new file mode 100644
index 00000000..4a32f2a3
--- /dev/null
+++ b/docs/benchmarks/artifacts/opmem-external-lanes/telemetry-zep-full.json
@@ -0,0 +1,247 @@
+{
+  "lane": "zep",
+  "usage": {
+    "chat_input_tokens": 0,
+    "chat_output_tokens": 0,
+    "embed_tokens": 0,
+    "estimated_usd": 0.0,
+    "http_requests": 47
+  },
+  "events": [
+    {
+      "op": "remember",
+      "status": 200,
+      "body_preview": "zep thread.add_messages"
+    },
+    {
+      "op": "recall",
+      "status": 200,
+      "body_preview": "[]"
+    },
+    {
+      "op": "revise",
+      "status": 200,
+      "body_preview": "zep thread.add_messages revise"
+    },
+    {
+      "op": "recall",
+      "status": 200,
+      "body_preview": "[{'id': 'zep-msg-1', 'content': 'Prefers Ruby for scripting'}]"
+    },
+    {
+      "op": "remember",
+      "status": 200,
+      "body_preview": "zep thread.add_messages"
+    },
+    {
+      "op": "revise",
+      "status": 200,
+      "body_preview": "zep thread.add_messages revise"
+    },
+    {
+      "op": "remember",
+      "status": 200,
+      "body_preview": "zep thread.add_messages"
+    },
+    {
+      "op": "recall",
+      "status": 200,
+      "body_preview": "[{'id': 'zep-msg-1', 'content': 'Prefers Ruby for scripting'}]"
+    },
+    {
+      "op": "remember",
+      "status": 200,
+      "body_preview": "zep thread.add_messages"
+    },
+    {
+      "op": "revise",
+      "status": 200,
+      "body_preview": "zep thread.add_messages revise"
+    },
+    {
+      "op": "recall",
+      "status": 200,
+      "body_preview": "[{'id': 'zep-msg-1', 'content': 'Prefers Ruby for scripting'}]"
+    },
+    {
+      "op": "remember",
+      "status": 200,
+      "body_preview": "zep thread.add_messages"
+    },
+    {
+      "op": "remember",
+      "status": 200,
+      "body_preview": "zep thread.add_messages"
+    },
+    {
+      "op": "recall",
+      "status": 200,
+      "body_preview": "[{'id': 'zep-msg-0', 'content': 'The wifi network name is FalconNet.'}]"
+    },
+    {
+      "op": "remember",
+      "status": 200,
+      "body_preview": "zep thread.add_messages"
+    },
+    {
+      "op": "recall",
+      "status": 200,
+      "body_preview": "[{'id': 'zep-msg-0', 'content': 'My name is Alice Chen.'}]"
+    },
+    {
+      "op": "recall",
+      "status": 200,
+      "body_preview": "[{'id': 'zep-msg-0', 'content': 'My name is Alice Chen.'}]"
+    },
+    {
+      "op": "remember",
+      "status": 200,
+      "body_preview": "zep thread.add_messages"
+    },
+    {
+      "op": "recall",
+      "status": 200,
+      "body_preview": "[{'id': 'zep-msg-0', 'content': 'The wifi network name is FalconNet.'}]"
+    },
+    {
+      "op": "recall",
+      "status": 200,
+      "body_preview": "[{'id': 'zep-msg-0', 'content': 'The wifi network name is FalconNet.'}]"
+    },
+    {
+      "op": "remember",
+      "status": 200,
+      "body_preview": "zep thread.add_messages"
+    },
+    {
+      "op": "remember",
+      "status": 200,
+      "body_preview": "zep thread.add_messages"
+    },
+    {
+      "op": "forget",
+      "status": 200,
+      "body_preview": "zep forget noop for ['m-ee3e2e287b7c']"
+    },
+    {
+      "op": "recall",
+      "status": 200,
+      "body_preview": "[{'id': 'zep-msg-0', 'content': 'Never share the door code with vendors.'}]"
+    },
+    {
+      "op": "recall",
+      "status": 200,
+      "body_preview": "[{'id': 'zep-msg-0', 'content': 'Never share the door code with vendors.'}]"
+    },
+    {
+      "op": "remember",
+      "status": 200,
+      "body_preview": "zep thread.add_messages"
+    },
+    {
+      "op": "recall",
+      "status": 200,
+      "body_preview": "[{'id': 'zep-msg-0', 'content': 'Never share the door code with vendors.'}]"
+    },
+    {
+      "op": "forget",
+      "status": 200,
+      "body_preview": "zep forget noop for ['m-ec19f9d4b669']"
+    },
+    {
+      "op": "recall",
+      "status": 200,
+      "body_preview": "[{'id': 'zep-msg-0', 'content': 'Never share the door code with vendors.'}]"
+    },
+    {
+      "op": "remember",
+      "status": 200,
+      "body_preview": "zep thread.add_messages"
+    },
+    {
+      "op": "remember",
+      "status": 200,
+      "body_preview": "zep thread.add_messages"
+    },
+    {
+      "op": "forget",
+      "status": 200,
+      "body_preview": "zep forget noop for ['m-14ada541a648']"
+    },
+    {
+      "op": "recall",
+      "status": 200,
+      "body_preview": "[{'id': 'zep-msg-0', 'content': 'Never share the door code with vendors.'}]"
+    },
+    {
+      "op": "recall",
+      "status": 200,
+      "body_preview": "[{'id': 'zep-msg-1', 'content': 'The wifi network name is FalconNet.'}]"
+    },
+    {
+      "op": "remember",
+      "status": 200,
+      "body_preview": "zep thread.add_messages"
+    },
+    {
+      "op": "forget",
+      "status": 200,
+      "body_preview": "zep forget noop for ['m-33684a38ce3c']"
+    },
+    {
+      "op": "remember",
+      "status": 200,
+      "body_preview": "zep thread.add_messages"
+    },
+    {
+      "op": "recall",
+      "status": 200,
+      "body_preview": "[{'id': 'zep-msg-0', 'content': 'Never share the door code with vendors.'}, {'id': 'cda0240f-f63a-4614-a24e-87a105cab156', 'content': 'OpMem Lane advises never sharing the door code with vendors.'}]"
+    },
+    {
+      "op": "remember",
+      "status": 200,
+      "body_preview": "zep thread.add_messages"
+    },
+    {
+      "op": "remember",
+      "status": 200,
+      "body_preview": "zep thread.add_messages"
+    },
+    {
+      "op": "recall",
+      "status": 200,
+      "body_preview": "[{'id': 'zep-msg-0', 'content': 'The launch date is May 12.'}, {'id': 'zep-msg-1', 'content': 'The launch date is June 3.'}, {'id': 'b61a203c-30ff-4aeb-90de-7b7d2da26023', 'content': 'OpMem Lane stated that the launch date is May 12.'}]"
+    },
+    {
+      "op": "remember",
+      "status": 200,
+      "body_preview": "zep thread.add_messages"
+    },
+    {
+      "op": "remember",
+      "status": 200,
+      "body_preview": "zep thread.add_messages"
+    },
+    {
+      "op": "recall",
+      "status": 200,
+      "body_preview": "[]"
+    },
+    {
+      "op": "remember",
+      "status": 200,
+      "body_preview": "zep thread.add_messages"
+    },
+    {
+      "op": "remember",
+      "status": 200,
+      "body_preview": "zep thread.add_messages"
+    },
+    {
+      "op": "recall",
+      "status": 200,
+      "body_preview": "[{'id': 'zep-msg-0', 'content': 'The door code is 1111.'}, {'id': 'zep-msg-1', 'content': 'The door code is 2222.'}, {'id': '838b7748-642b-4b60-95e0-cc09bb174823', 'content': \"OpMem Lane's door code is 1111.\"}]"
+    }
+  ]
+}
diff --git a/docs/benchmarks/artifacts/opmem-external-lanes/telemetry-zep-smoke.json b/docs/benchmarks/artifacts/opmem-external-lanes/telemetry-zep-smoke.json
new file mode 100644
index 00000000..38ca5c29
--- /dev/null
+++ b/docs/benchmarks/artifacts/opmem-external-lanes/telemetry-zep-smoke.json
@@ -0,0 +1,27 @@
+{
+  "lane": "zep",
+  "usage": {
+    "chat_input_tokens": 0,
+    "chat_output_tokens": 0,
+    "embed_tokens": 0,
+    "estimated_usd": 0.0,
+    "http_requests": 3
+  },
+  "events": [
+    {
+      "op": "remember",
+      "status": 200,
+      "body_preview": "zep thread.add_messages"
+    },
+    {
+      "op": "remember",
+      "status": 200,
+      "body_preview": "zep thread.add_messages"
+    },
+    {
+      "op": "recall",
+      "status": 200,
+      "body_preview": "[{'id': 'zep-msg-0', 'content': 'The wifi network name is FalconNet.'}, {'id': '413e2761-df41-49c5-ba6d-d37aaa89fe70', 'content': \"OpMem Lane's wifi network is named FalconNet.\"}]"
+    }
+  ]
+}
diff --git a/docs/benchmarks/artifacts/opmem-external-lanes/zep-env-check-2026-09-23.json b/docs/benchmarks/artifacts/opmem-external-lanes/zep-env-check-2026-09-23.json
new file mode 100644
index 00000000..eaf17b27
--- /dev/null
+++ b/docs/benchmarks/artifacts/opmem-external-lanes/zep-env-check-2026-09-23.json
@@ -0,0 +1,12 @@
+{
+  "checked_utc": "2026-09-23T17:55:00Z",
+  "zep_api_key_in_environment": true,
+  "zep_api_key_in_injected_secret_names": true,
+  "free_tier_preflight": "ok",
+  "zep_sdk": "zep-cloud==3.28.0",
+  "paid_ceiling_usd": 0.0,
+  "smoke_run_id": "opmem-phase2-zep-smoke-final",
+  "full_run_id": "opmem-phase2-zep-full",
+  "full_score": "3/13",
+  "infrastructure_errors": 0
+}
diff --git a/docs/benchmarks/artifacts/opmem-external-lanes/zep-score-interpretation-opmem-phase2-zep-full.json b/docs/benchmarks/artifacts/opmem-external-lanes/zep-score-interpretation-opmem-phase2-zep-full.json
new file mode 100644
index 00000000..528e8de1
--- /dev/null
+++ b/docs/benchmarks/artifacts/opmem-external-lanes/zep-score-interpretation-opmem-phase2-zep-full.json
@@ -0,0 +1,56 @@
+{
+  "source_artifact": "live-full-opmem-phase2-zep-full.json",
+  "run_id": "opmem-phase2-zep-full",
+  "harness_raw": {
+    "tasks_run": 13,
+    "tasks_passed": 3,
+    "label": "3/13"
+  },
+  "forget_not_supported": {
+    "reason": "zep-cloud 3.28 thread/graph API: adapter has no delete path for OpMem forget steps (documented noop).",
+    "tasks": [
+      "iso03_forget_isolated",
+      "sup01_basic_forget",
+      "sup02_targeted_forget",
+      "sup03_durable_forget"
+    ],
+    "harness_outcomes_not_counted_as_failures": [
+      {
+        "task": "iso03_forget_isolated",
+        "passed": false,
+        "failures": ["step3(recall): expected at most 0 results, got 1"]
+      },
+      {
+        "task": "sup01_basic_forget",
+        "passed": false,
+        "failures": ["step3(recall): expected at most 0 results, got 1"]
+      },
+      {
+        "task": "sup02_targeted_forget",
+        "passed": false,
+        "failures": ["step3(recall): a result contains excluded 'door code'"]
+      },
+      {
+        "task": "sup03_durable_forget",
+        "passed": false,
+        "failures": ["step3(recall): expected at most 0 results, got 2"]
+      }
+    ]
+  },
+  "scored_supported_tasks": {
+    "supported_task_count": 9,
+    "passed": 3,
+    "label": "3/9",
+    "passed_tasks": ["cor02_correction_stickiness", "cor03_revised_retrievable", "dup01_idempotent_remember"],
+    "failed_supported_tasks": [
+      "cor01_basic_revision",
+      "iso01_subject_isolation",
+      "iso02_tenant_isolation",
+      "upd01_stale_fact",
+      "upd02_preference_change",
+      "upd03_state_supersession"
+    ]
+  },
+  "publication_label": "3/9*",
+  "footnote": "Asterisk: four forget fixtures not supported on Zep lane; raw harness score remains 3/13 on opmem-phase2-zep-full."
+}
diff --git a/docs/benchmarks/artifacts/opmem-manuscript-20260918.md b/docs/benchmarks/artifacts/opmem-manuscript-20260918.md
new file mode 100644
index 00000000..eabcb317
--- /dev/null
+++ b/docs/benchmarks/artifacts/opmem-manuscript-20260918.md
@@ -0,0 +1,47 @@
+# OpMem manuscript pin — 2026-09-18
+
+**Git SHA:** `20533f2ca83e02265663d5984b49b13de8499231`  
+**Protocol:** [PUBLICATION_READINESS.md](../../research/opmem/PUBLICATION_READINESS.md)  
+**PDF:** [main.pdf](../../research/opmem/paper/main.pdf) (`make -C docs/research/opmem/paper pdf`)
+
+## Primary three-system run (embedded API + Mem0 Platform)
+
+| System | Overall |
+| --- | ---: |
+| Brainy (`/memories/search`) | **13/13** |
+| Verbatim baseline | **10/13** |
+| Mem0 Platform | **10/13** |
+
+Mem0 Platform: **completed** at negligible API volume (~47 scripted steps). Paid vendor runs elsewhere require a cost estimate and explicit approval first (not a categorical spend ban).
+
+Machine-readable: `opmem-pin-20260918.json` (copy of `opmem-manuscript-20533f2/opmem-primary.json`).
+
+## Lane ablations (isolated run)
+
+| Configuration | Overall |
+| --- | ---: |
+| `brainy` (search + correct) | **13/13** |
+| `brainy-recall` (`POST /recall`) | **9/13** |
+| `brainy-supersede` (`/supersede` revise) | **11/13** |
+
+Artifact: `opmem-manuscript-20533f2/opmem-lane-ablation-isolated.json`.  
+`brainy-recall` correction tasks hit HTTP 409 on `/correct` in this harness (documented threat).
+
+## Stability (5× repeat)
+
+Valid runs only (`infrastructure_errors=0`): **2/5** at **13/13**.  
+Runs reporting **9/10** with **3 infrastructure errors** are **invalid/incomplete** (errored tasks dropped from denominator—not comparable to 13/13). See `opmem-stability.json` and manuscript Section 6.
+
+## Reproduce
+
+```bash
+./scripts/run-opmem-manuscript.sh docs/benchmarks/artifacts/opmem-manuscript-$(git rev-parse --short HEAD)
+python3 scripts/generate-opmem-paper-tex.py
+make -C docs/research/opmem/paper pdf
+```
+
+## Pending external systems (blocked)
+
+Mem0 OSS, Zep, Letta, LangMem: **no OpMem adapter, credentials, or line-item cost worksheet** yet. Cost-estimation plan: manuscript Section 6 (`sec:cost-plan`) — 47 steps × mapped vendor billing units per system before approval.
+
+- v1 ~30 task expansion (separate track)
diff --git a/docs/benchmarks/artifacts/opmem-manuscript-20260918.pdf b/docs/benchmarks/artifacts/opmem-manuscript-20260918.pdf
new file mode 100644
index 00000000..d8e78fb2
Binary files /dev/null and b/docs/benchmarks/artifacts/opmem-manuscript-20260918.pdf differ
diff --git a/docs/benchmarks/artifacts/opmem-manuscript-20533f2/commands.log b/docs/benchmarks/artifacts/opmem-manuscript-20533f2/commands.log
new file mode 100644
index 00000000..ca1237b8
--- /dev/null
+++ b/docs/benchmarks/artifacts/opmem-manuscript-20533f2/commands.log
@@ -0,0 +1,5 @@
+git_sha=20533f2ca83e02265663d5984b49b13de8499231
+date_utc=2026-09-18T23:06:39Z
+go test ./internal/api/ -run TestOpMemBenchmarkAgainstHTTPServer -count=1
+OPMEM_MANUSCRIPT_OUT=docs/benchmarks/artifacts/opmem-manuscript-20533f2 go test ./internal/api/ -run TestOpMemPaperExperiments -count=1
+mem0 run: python3 evals/run_opmem.py --systems mem0 --json-out docs/benchmarks/artifacts/opmem-manuscript-20533f2/opmem-mem0-rerun.json
diff --git a/docs/benchmarks/artifacts/opmem-manuscript-20533f2/go-test-opmem.log b/docs/benchmarks/artifacts/opmem-manuscript-20533f2/go-test-opmem.log
new file mode 100644
index 00000000..c96079c6
--- /dev/null
+++ b/docs/benchmarks/artifacts/opmem-manuscript-20533f2/go-test-opmem.log
@@ -0,0 +1 @@
+ok  	brainy/internal/api	2.552s
diff --git a/docs/benchmarks/artifacts/opmem-manuscript-20533f2/go-test-paper.log b/docs/benchmarks/artifacts/opmem-manuscript-20533f2/go-test-paper.log
new file mode 100644
index 00000000..ffcb14f3
--- /dev/null
+++ b/docs/benchmarks/artifacts/opmem-manuscript-20533f2/go-test-paper.log
@@ -0,0 +1 @@
+ok  	brainy/internal/api	261.706s
diff --git a/docs/benchmarks/artifacts/opmem-manuscript-20533f2/mem0-rerun.log b/docs/benchmarks/artifacts/opmem-manuscript-20533f2/mem0-rerun.log
new file mode 100644
index 00000000..877e0bc9
--- /dev/null
+++ b/docs/benchmarks/artifacts/opmem-manuscript-20533f2/mem0-rerun.log
@@ -0,0 +1,157 @@
+{
+  "benchmark": "opmem-v0",
+  "systems": [
+    "mem0"
+  ],
+  "summary": {
+    "mem0": {
+      "overall": "10/13",
+      "by_category": {
+        "correction": "2/3",
+        "idempotency": "1/1",
+        "isolation": "3/3",
+        "staleness": "2/3",
+        "suppression": "2/3"
+      }
+    }
+  },
+  "infrastructure_errors": 0,
+  "results": [
+    {
+      "task": "cor01_basic_revision",
+      "category": "correction",
+      "systems": {
+        "mem0": {
+          "passed": true,
+          "failures": []
+        }
+      }
+    },
+    {
+      "task": "cor02_correction_stickiness",
+      "category": "correction",
+      "systems": {
+        "mem0": {
+          "passed": false,
+          "failures": [
+            "step3(recall): top result does not contain 'ruby'"
+          ]
+        }
+      }
+    },
+    {
+      "task": "cor03_revised_retrievable",
+      "category": "correction",
+      "systems": {
+        "mem0": {
+          "passed": true,
+          "failures": []
+        }
+      }
+    },
+    {
+      "task": "dup01_idempotent_remember",
+      "category": "idempotency",
+      "systems": {
+        "mem0": {
+          "passed": true,
+          "failures": []
+        }
+      }
+    },
+    {
+      "task": "iso01_subject_isolation",
+      "category": "isolation",
+      "systems": {
+        "mem0": {
+          "passed": true,
+          "failures": []
+        }
+      }
+    },
+    {
+      "task": "iso02_tenant_isolation",
+      "category": "isolation",
+      "systems": {
+        "mem0": {
+          "passed": true,
+          "failures": []
+        }
+      }
+    },
+    {
+      "task": "iso03_forget_isolated",
+      "category": "isolation",
+      "systems": {
+        "mem0": {
+          "passed": true,
+          "failures": []
+        }
+      }
+    },
+    {
+      "task": "sup01_basic_forget",
+      "category": "suppression",
+      "systems": {
+        "mem0": {
+          "passed": true,
+          "failures": []
+        }
+      }
+    },
+    {
+      "task": "sup02_targeted_forget",
+      "category": "suppression",
+      "systems": {
+        "mem0": {
+          "passed": true,
+          "failures": []
+        }
+      }
+    },
+    {
+      "task": "sup03_durable_forget",
+      "category": "suppression",
+      "systems": {
+        "mem0": {
+          "passed": false,
+          "failures": [
+            "step3(recall): expected at most 0 results, got 1"
+          ]
+        }
+      }
+    },
+    {
+      "task": "upd01_stale_fact",
+      "category": "staleness",
+      "systems": {
+        "mem0": {
+          "passed": true,
+          "failures": []
+        }
+      }
+    },
+    {
+      "task": "upd02_preference_change",
+      "category": "staleness",
+      "systems": {
+        "mem0": {
+          "passed": false,
+          "failures": [
+            "step2(recall): top result does not contain 'sms'"
+          ]
+        }
+      }
+    },
+    {
+      "task": "upd03_state_supersession",
+      "category": "staleness",
+      "systems": {
+        "mem0": {
+          "passed": true,
+          "failures": []
+        }
+      }
+    }
+  ]
+}
diff --git a/docs/benchmarks/artifacts/opmem-manuscript-20533f2/opmem-lane-ablation-isolated.json b/docs/benchmarks/artifacts/opmem-manuscript-20533f2/opmem-lane-ablation-isolated.json
new file mode 100644
index 00000000..a098cbb1
--- /dev/null
+++ b/docs/benchmarks/artifacts/opmem-manuscript-20533f2/opmem-lane-ablation-isolated.json
@@ -0,0 +1,282 @@
+{
+  "benchmark": "opmem-v0",
+  "systems": [
+    "brainy",
+    "brainy-recall",
+    "brainy-supersede"
+  ],
+  "summary": {
+    "brainy": {
+      "overall": "13/13",
+      "by_category": {
+        "correction": "3/3",
+        "idempotency": "1/1",
+        "isolation": "3/3",
+        "staleness": "3/3",
+        "suppression": "3/3"
+      }
+    },
+    "brainy-recall": {
+      "overall": "9/13",
+      "by_category": {
+        "correction": "0/3",
+        "idempotency": "1/1",
+        "isolation": "3/3",
+        "staleness": "3/3",
+        "suppression": "2/3"
+      }
+    },
+    "brainy-supersede": {
+      "overall": "11/13",
+      "by_category": {
+        "correction": "2/3",
+        "idempotency": "1/1",
+        "isolation": "3/3",
+        "staleness": "3/3",
+        "suppression": "2/3"
+      }
+    }
+  },
+  "infrastructure_errors": 3,
+  "results": [
+    {
+      "task": "cor01_basic_revision",
+      "category": "correction",
+      "systems": {
+        "brainy": {
+          "passed": true,
+          "failures": []
+        },
+        "brainy-recall": {
+          "error": "HTTPError: HTTP Error 409: Conflict"
+        },
+        "brainy-supersede": {
+          "passed": false,
+          "failures": [
+            "step1(recall): top result does not contain 'python'"
+          ]
+        }
+      }
+    },
+    {
+      "task": "cor02_correction_stickiness",
+      "category": "correction",
+      "systems": {
+        "brainy": {
+          "passed": true,
+          "failures": []
+        },
+        "brainy-recall": {
+          "error": "HTTPError: HTTP Error 409: Conflict"
+        },
+        "brainy-supersede": {
+          "passed": true,
+          "failures": []
+        }
+      }
+    },
+    {
+      "task": "cor03_revised_retrievable",
+      "category": "correction",
+      "systems": {
+        "brainy": {
+          "passed": true,
+          "failures": []
+        },
+        "brainy-recall": {
+          "error": "HTTPError: HTTP Error 409: Conflict"
+        },
+        "brainy-supersede": {
+          "passed": true,
+          "failures": []
+        }
+      }
+    },
+    {
+      "task": "dup01_idempotent_remember",
+      "category": "idempotency",
+      "systems": {
+        "brainy": {
+          "passed": true,
+          "failures": []
+        },
+        "brainy-recall": {
+          "passed": true,
+          "failures": []
+        },
+        "brainy-supersede": {
+          "passed": true,
+          "failures": []
+        }
+      }
+    },
+    {
+      "task": "iso01_subject_isolation",
+      "category": "isolation",
+      "systems": {
+        "brainy": {
+          "passed": true,
+          "failures": []
+        },
+        "brainy-recall": {
+          "passed": true,
+          "failures": []
+        },
+        "brainy-supersede": {
+          "passed": true,
+          "failures": []
+        }
+      }
+    },
+    {
+      "task": "iso02_tenant_isolation",
+      "category": "isolation",
+      "systems": {
+        "brainy": {
+          "passed": true,
+          "failures": []
+        },
+        "brainy-recall": {
+          "passed": true,
+          "failures": []
+        },
+        "brainy-supersede": {
+          "passed": true,
+          "failures": []
+        }
+      }
+    },
+    {
+      "task": "iso03_forget_isolated",
+      "category": "isolation",
+      "systems": {
+        "brainy": {
+          "passed": true,
+          "failures": []
+        },
+        "brainy-recall": {
+          "passed": true,
+          "failures": []
+        },
+        "brainy-supersede": {
+          "passed": true,
+          "failures": []
+        }
+      }
+    },
+    {
+      "task": "sup01_basic_forget",
+      "category": "suppression",
+      "systems": {
+        "brainy": {
+          "passed": true,
+          "failures": []
+        },
+        "brainy-recall": {
+          "passed": false,
+          "failures": [
+            "step1(recall): expected at least 1 results, got 0",
+            "step1(recall): top result does not contain 'door code'"
+          ]
+        },
+        "brainy-supersede": {
+          "passed": false,
+          "failures": [
+            "step1(recall): expected at least 1 results, got 0",
+            "step1(recall): top result does not contain 'door code'"
+          ]
+        }
+      }
+    },
+    {
+      "task": "sup02_targeted_forget",
+      "category": "suppression",
+      "systems": {
+        "brainy": {
+          "passed": true,
+          "failures": []
+        },
+        "brainy-recall": {
+          "passed": true,
+          "failures": []
+        },
+        "brainy-supersede": {
+          "passed": true,
+          "failures": []
+        }
+      }
+    },
+    {
+      "task": "sup03_durable_forget",
+      "category": "suppression",
+      "systems": {
+        "brainy": {
+          "passed": true,
+          "failures": []
+        },
+        "brainy-recall": {
+          "passed": true,
+          "failures": []
+        },
+        "brainy-supersede": {
+          "passed": true,
+          "failures": []
+        }
+      }
+    },
+    {
+      "task": "upd01_stale_fact",
+      "category": "staleness",
+      "systems": {
+        "brainy": {
+          "passed": true,
+          "failures": []
+        },
+        "brainy-recall": {
+          "passed": true,
+          "failures": []
+        },
+        "brainy-supersede": {
+          "passed": true,
+          "failures": []
+        }
+      }
+    },
+    {
+      "task": "upd02_preference_change",
+      "category": "staleness",
+      "systems": {
+        "brainy": {
+          "passed": true,
+          "failures": []
+        },
+        "brainy-recall": {
+          "passed": true,
+          "failures": []
+        },
+        "brainy-supersede": {
+          "passed": true,
+          "failures": []
+        }
+      }
+    },
+    {
+      "task": "upd03_state_supersession",
+      "category": "staleness",
+      "systems": {
+        "brainy": {
+          "passed": true,
+          "failures": []
+        },
+        "brainy-recall": {
+          "passed": true,
+          "failures": []
+        },
+        "brainy-supersede": {
+          "passed": true,
+          "failures": []
+        }
+      }
+    }
+  ]
+}
diff --git a/docs/benchmarks/artifacts/opmem-manuscript-20533f2/opmem-lane-ablation.json b/docs/benchmarks/artifacts/opmem-manuscript-20533f2/opmem-lane-ablation.json
new file mode 100644
index 00000000..0013bfb8
--- /dev/null
+++ b/docs/benchmarks/artifacts/opmem-manuscript-20533f2/opmem-lane-ablation.json
@@ -0,0 +1,290 @@
+{
+  "benchmark": "opmem-v0",
+  "systems": [
+    "brainy",
+    "brainy-recall",
+    "brainy-supersede"
+  ],
+  "summary": {
+    "brainy": {
+      "overall": "9/10",
+      "by_category": {
+        "idempotency": "1/1",
+        "isolation": "3/3",
+        "staleness": "3/3",
+        "suppression": "2/3"
+      }
+    },
+    "brainy-recall": {
+      "overall": "9/10",
+      "by_category": {
+        "idempotency": "1/1",
+        "isolation": "3/3",
+        "staleness": "3/3",
+        "suppression": "2/3"
+      }
+    },
+    "brainy-supersede": {
+      "overall": "11/13",
+      "by_category": {
+        "correction": "2/3",
+        "idempotency": "1/1",
+        "isolation": "3/3",
+        "staleness": "3/3",
+        "suppression": "2/3"
+      }
+    }
+  },
+  "infrastructure_errors": 6,
+  "results": [
+    {
+      "task": "cor01_basic_revision",
+      "category": "correction",
+      "systems": {
+        "brainy": {
+          "error": "HTTPError: HTTP Error 409: Conflict"
+        },
+        "brainy-recall": {
+          "error": "HTTPError: HTTP Error 409: Conflict"
+        },
+        "brainy-supersede": {
+          "passed": false,
+          "failures": [
+            "step1(recall): top result does not contain 'python'"
+          ]
+        }
+      }
+    },
+    {
+      "task": "cor02_correction_stickiness",
+      "category": "correction",
+      "systems": {
+        "brainy": {
+          "error": "HTTPError: HTTP Error 409: Conflict"
+        },
+        "brainy-recall": {
+          "error": "HTTPError: HTTP Error 409: Conflict"
+        },
+        "brainy-supersede": {
+          "passed": true,
+          "failures": []
+        }
+      }
+    },
+    {
+      "task": "cor03_revised_retrievable",
+      "category": "correction",
+      "systems": {
+        "brainy": {
+          "error": "HTTPError: HTTP Error 409: Conflict"
+        },
+        "brainy-recall": {
+          "error": "HTTPError: HTTP Error 409: Conflict"
+        },
+        "brainy-supersede": {
+          "passed": true,
+          "failures": []
+        }
+      }
+    },
+    {
+      "task": "dup01_idempotent_remember",
+      "category": "idempotency",
+      "systems": {
+        "brainy": {
+          "passed": true,
+          "failures": []
+        },
+        "brainy-recall": {
+          "passed": true,
+          "failures": []
+        },
+        "brainy-supersede": {
+          "passed": true,
+          "failures": []
+        }
+      }
+    },
+    {
+      "task": "iso01_subject_isolation",
+      "category": "isolation",
+      "systems": {
+        "brainy": {
+          "passed": true,
+          "failures": []
+        },
+        "brainy-recall": {
+          "passed": true,
+          "failures": []
+        },
+        "brainy-supersede": {
+          "passed": true,
+          "failures": []
+        }
+      }
+    },
+    {
+      "task": "iso02_tenant_isolation",
+      "category": "isolation",
+      "systems": {
+        "brainy": {
+          "passed": true,
+          "failures": []
+        },
+        "brainy-recall": {
+          "passed": true,
+          "failures": []
+        },
+        "brainy-supersede": {
+          "passed": true,
+          "failures": []
+        }
+      }
+    },
+    {
+      "task": "iso03_forget_isolated",
+      "category": "isolation",
+      "systems": {
+        "brainy": {
+          "passed": true,
+          "failures": []
+        },
+        "brainy-recall": {
+          "passed": true,
+          "failures": []
+        },
+        "brainy-supersede": {
+          "passed": true,
+          "failures": []
+        }
+      }
+    },
+    {
+      "task": "sup01_basic_forget",
+      "category": "suppression",
+      "systems": {
+        "brainy": {
+          "passed": false,
+          "failures": [
+            "step1(recall): expected at least 1 results, got 0",
+            "step1(recall): top result does not contain 'door code'"
+          ]
+        },
+        "brainy-recall": {
+          "passed": false,
+          "failures": [
+            "step1(recall): expected at least 1 results, got 0",
+            "step1(recall): top result does not contain 'door code'"
+          ]
+        },
+        "brainy-supersede": {
+          "passed": false,
+          "failures": [
+            "step1(recall): expected at least 1 results, got 0",
+            "step1(recall): top result does not contain 'door code'"
+          ]
+        }
+      }
+    },
+    {
+      "task": "sup02_targeted_forget",
+      "category": "suppression",
+      "systems": {
+        "brainy": {
+          "passed": true,
+          "failures": []
+        },
+        "brainy-recall": {
+          "passed": true,
+          "failures": []
+        },
+        "brainy-supersede": {
+          "passed": true,
+          "failures": []
+        }
+      }
+    },
+    {
+      "task": "sup03_durable_forget",
+      "category": "suppression",
+      "systems": {
+        "brainy": {
+          "passed": true,
+          "failures": []
+        },
+        "brainy-recall": {
+          "passed": true,
+          "failures": []
+        },
+        "brainy-supersede": {
+          "passed": true,
+          "failures": []
+        }
+      }
+    },
+    {
+      "task": "upd01_stale_fact",
+      "category": "staleness",
+      "systems": {
+        "brainy": {
+          "passed": true,
+          "failures": []
+        },
+        "brainy-recall": {
+          "passed": true,
+          "failures": []
+        },
+        "brainy-supersede": {
+          "passed": true,
+          "failures": []
+        }
+      }
+    },
+    {
+      "task": "upd02_preference_change",
+      "category": "staleness",
+      "systems": {
+        "brainy": {
+          "passed": true,
+          "failures": []
+        },
+        "brainy-recall": {
+          "passed": true,
+          "failures": []
+        },
+        "brainy-supersede": {
+          "passed": true,
+          "failures": []
+        }
+      }
+    },
+    {
+      "task": "upd03_state_supersession",
+      "category": "staleness",
+      "systems": {
+        "brainy": {
+          "passed": true,
+          "failures": []
+        },
+        "brainy-recall": {
+          "passed": true,
+          "failures": []
+        },
+        "brainy-supersede": {
+          "passed": true,
+          "failures": []
+        }
+      }
+    }
+  ],
+  "manifest": {
+    "benchmark": "opmem-v0",
+    "git_sha": "20533f2ca83e02265663d5984b49b13de8499231",
+    "run_utc": "2026-09-18T23:12:29Z",
+    "fixture_dir": "fixtures/opmem",
+    "task_count": 13,
+    "base_url": "http://127.0.0.1:38319",
+    "protocol": "docs/research/opmem/PUBLICATION_READINESS.md",
+    "scoring": "substring assertions on ranked recall results"
+  }
+}
diff --git a/docs/benchmarks/artifacts/opmem-manuscript-20533f2/opmem-mechanism.json b/docs/benchmarks/artifacts/opmem-manuscript-20533f2/opmem-mechanism.json
new file mode 100644
index 00000000..1c5bfcbb
--- /dev/null
+++ b/docs/benchmarks/artifacts/opmem-manuscript-20533f2/opmem-mechanism.json
@@ -0,0 +1,37 @@
+{
+  "manifest": {
+    "benchmark": "opmem-v0",
+    "git_sha": "20533f2ca83e02265663d5984b49b13de8499231",
+    "run_utc": "2026-09-18T23:12:29Z",
+    "fixture_dir": "fixtures/opmem",
+    "task_count": 13,
+    "base_url": "http://127.0.0.1:38319",
+    "protocol": "docs/research/opmem/PUBLICATION_READINESS.md",
+    "scoring": "substring assertions on ranked recall results"
+  },
+  "by_category": {
+    "verbatim": {
+      "correction": "2/3",
+      "idempotency": "0/1",
+      "isolation": "3/3",
+      "staleness": "3/3",
+      "suppression": "2/3"
+    },
+    "brainy": {
+      "correction": "3/3",
+      "idempotency": "1/1",
+      "isolation": "3/3",
+      "staleness": "3/3",
+      "suppression": "3/3"
+    },
+    "mem0": {
+      "correction": "2/3",
+      "idempotency": "1/1",
+      "isolation": "3/3",
+      "staleness": "2/3",
+      "suppression": "2/3"
+    },
+    "_note": "verbatim is the rank-only baseline (no correct/suppress lifecycle); correction/suppression/staleness categories map to revise/forget semantics."
+  },
+  "verbatim_overall": "10/13"
+}
diff --git a/docs/benchmarks/artifacts/opmem-manuscript-20533f2/opmem-mem0-rerun.json b/docs/benchmarks/artifacts/opmem-manuscript-20533f2/opmem-mem0-rerun.json
new file mode 100644
index 00000000..877e0bc9
--- /dev/null
+++ b/docs/benchmarks/artifacts/opmem-manuscript-20533f2/opmem-mem0-rerun.json
@@ -0,0 +1,157 @@
+{
+  "benchmark": "opmem-v0",
+  "systems": [
+    "mem0"
+  ],
+  "summary": {
+    "mem0": {
+      "overall": "10/13",
+      "by_category": {
+        "correction": "2/3",
+        "idempotency": "1/1",
+        "isolation": "3/3",
+        "staleness": "2/3",
+        "suppression": "2/3"
+      }
+    }
+  },
+  "infrastructure_errors": 0,
+  "results": [
+    {
+      "task": "cor01_basic_revision",
+      "category": "correction",
+      "systems": {
+        "mem0": {
+          "passed": true,
+          "failures": []
+        }
+      }
+    },
+    {
+      "task": "cor02_correction_stickiness",
+      "category": "correction",
+      "systems": {
+        "mem0": {
+          "passed": false,
+          "failures": [
+            "step3(recall): top result does not contain 'ruby'"
+          ]
+        }
+      }
+    },
+    {
+      "task": "cor03_revised_retrievable",
+      "category": "correction",
+      "systems": {
+        "mem0": {
+          "passed": true,
+          "failures": []
+        }
+      }
+    },
+    {
+      "task": "dup01_idempotent_remember",
+      "category": "idempotency",
+      "systems": {
+        "mem0": {
+          "passed": true,
+          "failures": []
+        }
+      }
+    },
+    {
+      "task": "iso01_subject_isolation",
+      "category": "isolation",
+      "systems": {
+        "mem0": {
+          "passed": true,
+          "failures": []
+        }
+      }
+    },
+    {
+      "task": "iso02_tenant_isolation",
+      "category": "isolation",
+      "systems": {
+        "mem0": {
+          "passed": true,
+          "failures": []
+        }
+      }
+    },
+    {
+      "task": "iso03_forget_isolated",
+      "category": "isolation",
+      "systems": {
+        "mem0": {
+          "passed": true,
+          "failures": []
+        }
+      }
+    },
+    {
+      "task": "sup01_basic_forget",
+      "category": "suppression",
+      "systems": {
+        "mem0": {
+          "passed": true,
+          "failures": []
+        }
+      }
+    },
+    {
+      "task": "sup02_targeted_forget",
+      "category": "suppression",
+      "systems": {
+        "mem0": {
+          "passed": true,
+          "failures": []
+        }
+      }
+    },
+    {
+      "task": "sup03_durable_forget",
+      "category": "suppression",
+      "systems": {
+        "mem0": {
+          "passed": false,
+          "failures": [
+            "step3(recall): expected at most 0 results, got 1"
+          ]
+        }
+      }
+    },
+    {
+      "task": "upd01_stale_fact",
+      "category": "staleness",
+      "systems": {
+        "mem0": {
+          "passed": true,
+          "failures": []
+        }
+      }
+    },
+    {
+      "task": "upd02_preference_change",
+      "category": "staleness",
+      "systems": {
+        "mem0": {
+          "passed": false,
+          "failures": [
+            "step2(recall): top result does not contain 'sms'"
+          ]
+        }
+      }
+    },
+    {
+      "task": "upd03_state_supersession",
+      "category": "staleness",
+      "systems": {
+        "mem0": {
+          "passed": true,
+          "failures": []
+        }
+      }
+    }
+  ]
+}
diff --git a/docs/benchmarks/artifacts/opmem-manuscript-20533f2/opmem-primary.json b/docs/benchmarks/artifacts/opmem-manuscript-20533f2/opmem-primary.json
new file mode 100644
index 00000000..11550470
--- /dev/null
+++ b/docs/benchmarks/artifacts/opmem-manuscript-20533f2/opmem-primary.json
@@ -0,0 +1,300 @@
+{
+  "benchmark": "opmem-v0",
+  "systems": [
+    "verbatim",
+    "brainy",
+    "mem0"
+  ],
+  "summary": {
+    "verbatim": {
+      "overall": "10/13",
+      "by_category": {
+        "correction": "2/3",
+        "idempotency": "0/1",
+        "isolation": "3/3",
+        "staleness": "3/3",
+        "suppression": "2/3"
+      }
+    },
+    "brainy": {
+      "overall": "13/13",
+      "by_category": {
+        "correction": "3/3",
+        "idempotency": "1/1",
+        "isolation": "3/3",
+        "staleness": "3/3",
+        "suppression": "3/3"
+      }
+    },
+    "mem0": {
+      "overall": "10/13",
+      "by_category": {
+        "correction": "2/3",
+        "idempotency": "1/1",
+        "isolation": "3/3",
+        "staleness": "2/3",
+        "suppression": "2/3"
+      }
+    }
+  },
+  "infrastructure_errors": 0,
+  "results": [
+    {
+      "task": "cor01_basic_revision",
+      "category": "correction",
+      "systems": {
+        "verbatim": {
+          "passed": true,
+          "failures": []
+        },
+        "brainy": {
+          "passed": true,
+          "failures": []
+        },
+        "mem0": {
+          "passed": true,
+          "failures": []
+        }
+      }
+    },
+    {
+      "task": "cor02_correction_stickiness",
+      "category": "correction",
+      "systems": {
+        "verbatim": {
+          "passed": false,
+          "failures": [
+            "step3(recall): top result does not contain 'ruby'"
+          ]
+        },
+        "brainy": {
+          "passed": true,
+          "failures": []
+        },
+        "mem0": {
+          "passed": false,
+          "failures": [
+            "step3(recall): top result does not contain 'ruby'"
+          ]
+        }
+      }
+    },
+    {
+      "task": "cor03_revised_retrievable",
+      "category": "correction",
+      "systems": {
+        "verbatim": {
+          "passed": true,
+          "failures": []
+        },
+        "brainy": {
+          "passed": true,
+          "failures": []
+        },
+        "mem0": {
+          "passed": true,
+          "failures": []
+        }
+      }
+    },
+    {
+      "task": "dup01_idempotent_remember",
+      "category": "idempotency",
+      "systems": {
+        "verbatim": {
+          "passed": false,
+          "failures": [
+            "step2(recall): expected at most 1 results, got 2",
+            "step2(recall): duplicate result contents"
+          ]
+        },
+        "brainy": {
+          "passed": true,
+          "failures": []
+        },
+        "mem0": {
+          "passed": true,
+          "failures": []
+        }
+      }
+    },
+    {
+      "task": "iso01_subject_isolation",
+      "category": "isolation",
+      "systems": {
+        "verbatim": {
+          "passed": true,
+          "failures": []
+        },
+        "brainy": {
+          "passed": true,
+          "failures": []
+        },
+        "mem0": {
+          "passed": true,
+          "failures": []
+        }
+      }
+    },
+    {
+      "task": "iso02_tenant_isolation",
+      "category": "isolation",
+      "systems": {
+        "verbatim": {
+          "passed": true,
+          "failures": []
+        },
+        "brainy": {
+          "passed": true,
+          "failures": []
+        },
+        "mem0": {
+          "passed": true,
+          "failures": []
+        }
+      }
+    },
+    {
+      "task": "iso03_forget_isolated",
+      "category": "isolation",
+      "systems": {
+        "verbatim": {
+          "passed": true,
+          "failures": []
+        },
+        "brainy": {
+          "passed": true,
+          "failures": []
+        },
+        "mem0": {
+          "passed": true,
+          "failures": []
+        }
+      }
+    },
+    {
+      "task": "sup01_basic_forget",
+      "category": "suppression",
+      "systems": {
+        "verbatim": {
+          "passed": true,
+          "failures": []
+        },
+        "brainy": {
+          "passed": true,
+          "failures": []
+        },
+        "mem0": {
+          "passed": true,
+          "failures": []
+        }
+      }
+    },
+    {
+      "task": "sup02_targeted_forget",
+      "category": "suppression",
+      "systems": {
+        "verbatim": {
+          "passed": true,
+          "failures": []
+        },
+        "brainy": {
+          "passed": true,
+          "failures": []
+        },
+        "mem0": {
+          "passed": true,
+          "failures": []
+        }
+      }
+    },
+    {
+      "task": "sup03_durable_forget",
+      "category": "suppression",
+      "systems": {
+        "verbatim": {
+          "passed": false,
+          "failures": [
+            "step3(recall): expected at most 0 results, got 1"
+          ]
+        },
+        "brainy": {
+          "passed": true,
+          "failures": []
+        },
+        "mem0": {
+          "passed": false,
+          "failures": [
+            "step3(recall): expected at most 0 results, got 1"
+          ]
+        }
+      }
+    },
+    {
+      "task": "upd01_stale_fact",
+      "category": "staleness",
+      "systems": {
+        "verbatim": {
+          "passed": true,
+          "failures": []
+        },
+        "brainy": {
+          "passed": true,
+          "failures": []
+        },
+        "mem0": {
+          "passed": true,
+          "failures": []
+        }
+      }
+    },
+    {
+      "task": "upd02_preference_change",
+      "category": "staleness",
+      "systems": {
+        "verbatim": {
+          "passed": true,
+          "failures": []
+        },
+        "brainy": {
+          "passed": true,
+          "failures": []
+        },
+        "mem0": {
+          "passed": false,
+          "failures": [
+            "step2(recall): top result does not contain 'sms'"
+          ]
+        }
+      }
+    },
+    {
+      "task": "upd03_state_supersession",
+      "category": "staleness",
+      "systems": {
+        "verbatim": {
+          "passed": true,
+          "failures": []
+        },
+        "brainy": {
+          "passed": true,
+          "failures": []
+        },
+        "mem0": {
+          "passed": true,
+          "failures": []
+        }
+      }
+    }
+  ],
+  "manifest": {
+    "benchmark": "opmem-v0",
+    "git_sha": "20533f2ca83e02265663d5984b49b13de8499231",
+    "run_utc": "2026-09-18T23:12:29Z",
+    "fixture_dir": "fixtures/opmem",
+    "task_count": 13,
+    "base_url": "http://127.0.0.1:38319",
+    "protocol": "docs/research/opmem/PUBLICATION_READINESS.md",
+    "scoring": "substring assertions on ranked recall results"
+  }
+}
diff --git a/docs/benchmarks/artifacts/opmem-manuscript-20533f2/opmem-stability.json b/docs/benchmarks/artifacts/opmem-manuscript-20533f2/opmem-stability.json
new file mode 100644
index 00000000..8e742796
--- /dev/null
+++ b/docs/benchmarks/artifacts/opmem-manuscript-20533f2/opmem-stability.json
@@ -0,0 +1,48 @@
+{
+  "repeats": 5,
+  "system": "brainy",
+  "recall_lane": "search",
+  "revise_mode": "correct",
+  "runs": [
+    {
+      "repeat": 1,
+      "overall": "13/13",
+      "infrastructure_errors": 0
+    },
+    {
+      "repeat": 2,
+      "overall": "9/10",
+      "infrastructure_errors": 3
+    },
+    {
+      "repeat": 3,
+      "overall": "9/10",
+      "infrastructure_errors": 3
+    },
+    {
+      "repeat": 4,
+      "overall": "13/13",
+      "infrastructure_errors": 0
+    },
+    {
+      "repeat": 5,
+      "overall": "9/10",
+      "infrastructure_errors": 3
+    }
+  ],
+  "stable": false,
+  "unique_scores": [
+    "13/13",
+    "9/10"
+  ],
+  "manifest": {
+    "benchmark": "opmem-v0",
+    "git_sha": "20533f2ca83e02265663d5984b49b13de8499231",
+    "run_utc": "2026-09-18T23:12:29Z",
+    "fixture_dir": "fixtures/opmem",
+    "task_count": 13,
+    "base_url": "http://127.0.0.1:38319",
+    "protocol": "docs/research/opmem/PUBLICATION_READINESS.md",
+    "scoring": "substring assertions on ranked recall results"
+  }
+}
diff --git a/docs/benchmarks/artifacts/opmem-manuscript-20533f2/run-manifest.json b/docs/benchmarks/artifacts/opmem-manuscript-20533f2/run-manifest.json
new file mode 100644
index 00000000..bac4adb2
--- /dev/null
+++ b/docs/benchmarks/artifacts/opmem-manuscript-20533f2/run-manifest.json
@@ -0,0 +1,10 @@
+{
+  "benchmark": "opmem-v0",
+  "git_sha": "20533f2ca83e02265663d5984b49b13de8499231",
+  "run_utc": "2026-09-18T23:12:29Z",
+  "fixture_dir": "fixtures/opmem",
+  "task_count": 13,
+  "base_url": "http://127.0.0.1:38319",
+  "protocol": "docs/research/opmem/PUBLICATION_READINESS.md",
+  "scoring": "substring assertions on ranked recall results"
+}
diff --git a/docs/benchmarks/artifacts/opmem-manuscript-20533f2/run.log b/docs/benchmarks/artifacts/opmem-manuscript-20533f2/run.log
new file mode 100644
index 00000000..e942a6c1
--- /dev/null
+++ b/docs/benchmarks/artifacts/opmem-manuscript-20533f2/run.log
@@ -0,0 +1,5 @@
+git_sha=20533f2ca83e02265663d5984b49b13de8499231
+tasks=13
+primary={"verbatim": {"overall": "10/13", "by_category": {"correction": "2/3", "idempotency": "0/1", "isolation": "3/3", "staleness": "3/3", "suppression": "2/3"}}, "brainy": {"overall": "13/13", "by_category": {"correction": "3/3", "idempotency": "1/1", "isolation": "3/3", "staleness": "3/3", "suppression": "3/3"}}, "mem0": {"overall": "10/13", "by_category": {"correction": "2/3", "idempotency": "1/1", "isolation": "3/3", "staleness": "2/3", "suppression": "2/3"}}}
+lanes={"brainy": {"overall": "9/10", "by_category": {"idempotency": "1/1", "isolation": "3/3", "staleness": "3/3", "suppression": "2/3"}}, "brainy-recall": {"overall": "9/10", "by_category": {"idempotency": "1/1", "isolation": "3/3", "staleness": "3/3", "suppression": "2/3"}}, "brainy-supersede": {"overall": "11/13", "by_category": {"correction": "2/3", "idempotency": "1/1", "isolation": "3/3", "staleness": "3/3", "suppression": "2/3"}}}
+stability={"repeats": 5, "system": "brainy", "recall_lane": "search", "revise_mode": "correct", "runs": [{"repeat": 1, "overall": "13/13", "infrastructure_errors": 0}, {"repeat": 2, "overall": "9/10", "infrastructure_errors": 3}, {"repeat": 3, "overall": "9/10", "infrastructure_errors": 3}, {"repeat": 4, "overall": "13/13", "infrastructure_errors": 0}, {"repeat": 5, "overall": "9/10", "infrastructure_errors": 3}], "stable": false, "unique_scores": ["13/13", "9/10"], "manifest": {"benchmark": "opmem-v0", "git_sha": "20533f2ca83e02265663d5984b49b13de8499231", "run_utc": "2026-09-18T23:12:29Z", "fixture_dir": "fixtures/opmem", "task_count": 13, "base_url": "http://127.0.0.1:38319", "protocol": "docs/research/opmem/PUBLICATION_READINESS.md", "scoring": "substring assertions on ranked recall results"}}
diff --git a/docs/benchmarks/artifacts/opmem-pin-20260918.json b/docs/benchmarks/artifacts/opmem-pin-20260918.json
new file mode 100644
index 00000000..11550470
--- /dev/null
+++ b/docs/benchmarks/artifacts/opmem-pin-20260918.json
@@ -0,0 +1,300 @@
+{
+  "benchmark": "opmem-v0",
+  "systems": [
+    "verbatim",
+    "brainy",
+    "mem0"
+  ],
+  "summary": {
+    "verbatim": {
+      "overall": "10/13",
+      "by_category": {
+        "correction": "2/3",
+        "idempotency": "0/1",
+        "isolation": "3/3",
+        "staleness": "3/3",
+        "suppression": "2/3"
+      }
+    },
+    "brainy": {
+      "overall": "13/13",
+      "by_category": {
+        "correction": "3/3",
+        "idempotency": "1/1",
+        "isolation": "3/3",
+        "staleness": "3/3",
+        "suppression": "3/3"
+      }
+    },
+    "mem0": {
+      "overall": "10/13",
+      "by_category": {
+        "correction": "2/3",
+        "idempotency": "1/1",
+        "isolation": "3/3",
+        "staleness": "2/3",
+        "suppression": "2/3"
+      }
+    }
+  },
+  "infrastructure_errors": 0,
+  "results": [
+    {
+      "task": "cor01_basic_revision",
+      "category": "correction",
+      "systems": {
+        "verbatim": {
+          "passed": true,
+          "failures": []
+        },
+        "brainy": {
+          "passed": true,
+          "failures": []
+        },
+        "mem0": {
+          "passed": true,
+          "failures": []
+        }
+      }
+    },
+    {
+      "task": "cor02_correction_stickiness",
+      "category": "correction",
+      "systems": {
+        "verbatim": {
+          "passed": false,
+          "failures": [
+            "step3(recall): top result does not contain 'ruby'"
+          ]
+        },
+        "brainy": {
+          "passed": true,
+          "failures": []
+        },
+        "mem0": {
+          "passed": false,
+          "failures": [
+            "step3(recall): top result does not contain 'ruby'"
+          ]
+        }
+      }
+    },
+    {
+      "task": "cor03_revised_retrievable",
+      "category": "correction",
+      "systems": {
+        "verbatim": {
+          "passed": true,
+          "failures": []
+        },
+        "brainy": {
+          "passed": true,
+          "failures": []
+        },
+        "mem0": {
+          "passed": true,
+          "failures": []
+        }
+      }
+    },
+    {
+      "task": "dup01_idempotent_remember",
+      "category": "idempotency",
+      "systems": {
+        "verbatim": {
+          "passed": false,
+          "failures": [
+            "step2(recall): expected at most 1 results, got 2",
+            "step2(recall): duplicate result contents"
+          ]
+        },
+        "brainy": {
+          "passed": true,
+          "failures": []
+        },
+        "mem0": {
+          "passed": true,
+          "failures": []
+        }
+      }
+    },
+    {
+      "task": "iso01_subject_isolation",
+      "category": "isolation",
+      "systems": {
+        "verbatim": {
+          "passed": true,
+          "failures": []
+        },
+        "brainy": {
+          "passed": true,
+          "failures": []
+        },
+        "mem0": {
+          "passed": true,
+          "failures": []
+        }
+      }
+    },
+    {
+      "task": "iso02_tenant_isolation",
+      "category": "isolation",
+      "systems": {
+        "verbatim": {
+          "passed": true,
+          "failures": []
+        },
+        "brainy": {
+          "passed": true,
+          "failures": []
+        },
+        "mem0": {
+          "passed": true,
+          "failures": []
+        }
+      }
+    },
+    {
+      "task": "iso03_forget_isolated",
+      "category": "isolation",
+      "systems": {
+        "verbatim": {
+          "passed": true,
+          "failures": []
+        },
+        "brainy": {
+          "passed": true,
+          "failures": []
+        },
+        "mem0": {
+          "passed": true,
+          "failures": []
+        }
+      }
+    },
+    {
+      "task": "sup01_basic_forget",
+      "category": "suppression",
+      "systems": {
+        "verbatim": {
+          "passed": true,
+          "failures": []
+        },
+        "brainy": {
+          "passed": true,
+          "failures": []
+        },
+        "mem0": {
+          "passed": true,
+          "failures": []
+        }
+      }
+    },
+    {
+      "task": "sup02_targeted_forget",
+      "category": "suppression",
+      "systems": {
+        "verbatim": {
+          "passed": true,
+          "failures": []
+        },
+        "brainy": {
+          "passed": true,
+          "failures": []
+        },
+        "mem0": {
+          "passed": true,
+          "failures": []
+        }
+      }
+    },
+    {
+      "task": "sup03_durable_forget",
+      "category": "suppression",
+      "systems": {
+        "verbatim": {
+          "passed": false,
+          "failures": [
+            "step3(recall): expected at most 0 results, got 1"
+          ]
+        },
+        "brainy": {
+          "passed": true,
+          "failures": []
+        },
+        "mem0": {
+          "passed": false,
+          "failures": [
+            "step3(recall): expected at most 0 results, got 1"
+          ]
+        }
+      }
+    },
+    {
+      "task": "upd01_stale_fact",
+      "category": "staleness",
+      "systems": {
+        "verbatim": {
+          "passed": true,
+          "failures": []
+        },
+        "brainy": {
+          "passed": true,
+          "failures": []
+        },
+        "mem0": {
+          "passed": true,
+          "failures": []
+        }
+      }
+    },
+    {
+      "task": "upd02_preference_change",
+      "category": "staleness",
+      "systems": {
+        "verbatim": {
+          "passed": true,
+          "failures": []
+        },
+        "brainy": {
+          "passed": true,
+          "failures": []
+        },
+        "mem0": {
+          "passed": false,
+          "failures": [
+            "step2(recall): top result does not contain 'sms'"
+          ]
+        }
+      }
+    },
+    {
+      "task": "upd03_state_supersession",
+      "category": "staleness",
+      "systems": {
+        "verbatim": {
+          "passed": true,
+          "failures": []
+        },
+        "brainy": {
+          "passed": true,
+          "failures": []
+        },
+        "mem0": {
+          "passed": true,
+          "failures": []
+        }
+      }
+    }
+  ],
+  "manifest": {
+    "benchmark": "opmem-v0",
+    "git_sha": "20533f2ca83e02265663d5984b49b13de8499231",
+    "run_utc": "2026-09-18T23:12:29Z",
+    "fixture_dir": "fixtures/opmem",
+    "task_count": 13,
+    "base_url": "http://127.0.0.1:38319",
+    "protocol": "docs/research/opmem/PUBLICATION_READINESS.md",
+    "scoring": "substring assertions on ranked recall results"
+  }
+}
diff --git a/docs/research/master-plan.md b/docs/research/master-plan.md
index 765090e2..b342ace3 100644
--- a/docs/research/master-plan.md
+++ b/docs/research/master-plan.md
@@ -44,7 +44,7 @@ Staging verify: OpMem 13/13, support 3/3, `/recall` enumerate OK.
 
 | Axis | Brainy | Best competitor signal | Comparable? |
 | --- | --- | --- | --- |
-| OpMem (ops correctness, 12 tasks) | **12/12** | Mem0 platform 9/12 (measured, same fixtures) | Yes — our harness, both measured |
+| OpMem (ops correctness, 13 tasks) | **13/13** (`1b5ab3e` pin) | Mem0 Platform **10/13** (`2026-08-15`; July **9/12** was 12-task set) | Yes — our harness; Platform not OSS |
 | Marketing vertical (16 fixtures) | **15/16** empirical | Mem0 4/16 (measured; strict schema = intentional moat) | Yes, with declared policy |
 | LOCOMO smoke (1 convo × 30 Q, gpt-oss pin) | 19/30, MH 5/10 | Mem0 same-pin 12/30, MH 6/10 | Only vs our own same-pin Mem0 run |
 | LOCOMO full (10 convos, 1,540 Q) | **not run** (partial 3-convo old-stack: 31/90) | Mem0 platform 92.5; Hindsight 92.0; corrected Zep 75.1; Letta-filesystem 74.0; full-context ~73 | No — we have no comparable number |
diff --git a/docs/research/mem0-parity-gaps.md b/docs/research/mem0-parity-gaps.md
index d0fe488e..881deddf 100644
--- a/docs/research/mem0-parity-gaps.md
+++ b/docs/research/mem0-parity-gaps.md
@@ -11,7 +11,7 @@
 | Axis | Brainy | Mem0 (known) | Verdict |
 | --- | --- | --- | --- |
 | Thin-slice API parity | **4/4** | **4/4** | **Parity** |
-| Operational correctness (OpMem) | **12/12** | **9/12** | **Ahead** |
+| Operational correctness (OpMem) | **13/13** (`1b5ab3e` pin) | **10/13** Platform (`2026-08-15`) | **Ahead on frozen ops pin** (not LoCoMo) |
 | Vertical / governed memory | **16/16** marketing | No equivalent | **Ahead (category)** |
 | Conversational recall (LOCOMO) | **19/30** smoke (gpt-oss, 1×30) | ~92 platform (GPT, top-200, full suite) | **Behind — not comparable pins** |
 | Latency / tokens | p50 ~0.7–1.3s (staging) | Blog ~0.88s p50 @ top-200 | **Unknown fair pin** |
diff --git a/docs/research/opmem-spec.md b/docs/research/opmem-spec.md
index c7a21db6..99fe65f9 100644
--- a/docs/research/opmem-spec.md
+++ b/docs/research/opmem-spec.md
@@ -67,14 +67,14 @@ Recall assertions (all content matching is case-insensitive substring):
 | `must_exclude` | no result contains any of these substrings |
 | `unique_contents` | no two results share identical normalized content |
 
-### Task taxonomy (v0: 12 tasks)
+### Task taxonomy (v0: 13 tasks)
 
 | Category | Tasks | Probes |
 | --- | --- | --- |
 | suppression | `sup01` basic forget, `sup02` targeted forget, `sup03` durable forget | forget leaks; collateral damage; resurrection via re-ingestion |
 | correction | `cor01` basic revision, `cor02` correction stickiness, `cor03` revised retrievable | revision visibility; stale content winning after re-ingestion; retrieval by new terms |
 | isolation | `iso01` subject, `iso02` tenant, `iso03` forget isolation | cross-actor reads; cross-tenant reads; delete affecting a lookalike memory of another actor |
-| staleness | `upd01` stale fact, `upd02` preference change | latest version of a changed fact/preference must outrank the stale one |
+| staleness | `upd01` stale fact, `upd02` preference change, `upd03` state supersession | latest version of a changed fact/preference must outrank the stale one; restated state facts prefer newer value |
 | idempotency | `dup01` idempotent remember | duplicate ingestion polluting recall |
 
 Contents are deliberately domain-neutral (door codes, wifi names, launch dates,
@@ -130,7 +130,7 @@ CI runs the harness end-to-end via `TestOpMemBenchmarkAgainstHTTPServer`
 
 ## Roadmap to the paper
 
-- v0 (this): 12 tasks, 3 adapters (Brainy, verbatim baseline, Mem0), binary scoring.
+- v0 (this): 13 tasks (`upd03` added 2026-08), 3 adapters (Brainy, verbatim baseline, Mem0 Platform), binary scoring.
 - v1: add adapters for Zep, Letta, LangMem; expand to ~30 tasks including
   lifecycle/expiry (needs a neutral expiry attribute), batched/async ingestion
   visibility, and concurrent-write races.
diff --git a/docs/research/opmem/PUBLICATION_READINESS.md b/docs/research/opmem/PUBLICATION_READINESS.md
new file mode 100644
index 00000000..5d1bdf97
--- /dev/null
+++ b/docs/research/opmem/PUBLICATION_READINESS.md
@@ -0,0 +1,254 @@
+# OpMem publication readiness package
+
+**Status:** documentation only · **not a pin** · **no benchmark run in this pass**  
+**Inventory SHA:** `bbe55f7` (`origin/dev` / `origin/main` at package authoring)  
+**Authoring date:** 2026-09-18  
+
+This package consolidates repo evidence for Paper 1 ([paper-topics.md](../paper-topics.md): *Beyond recall: lifecycle, suppression, and correction correctness*). It does not claim new measurements.
+
+---
+
+## Ready vs missing (executive)
+
+| Area | Ready | Missing |
+| --- | --- | --- |
+| Task fixtures + runner | 13 JSON tasks, `evals/run_opmem.py`, Brainy/verbatim/Platform + **Phase 1** external lane stubs (`evals/opmem_external/`) | v1 ~30 tasks; **live** Zep/Letta/LangMem/Mem0 OSS runs (credentials + worksheet approval) |
+| Reproducible protocol | Documented below + [opmem-spec.md](../opmem-spec.md); CI spins embedded API | Pin file (`RunManifest`) for OpMem; SHA `bbe55f7` re-run of Brainy+Mem0 table |
+| Brainy 13/13 evidence | Markdown pin + per-task JSON for Brainy-only integrity pass | Full `result.json` with Brainy+verbatim+mem0 on one commit in git |
+| Mem0 10/13 evidence | Markdown pin with named failures | Per-task Mem0 JSON in git; Mem0 OSS run; API version drift check |
+| Manuscript | LaTeX + PDF at [paper/main.pdf](./paper/main.pdf) | arXiv submission; formal grammar; camera-ready figures |
+| Multi-system paper table | Brainy vs Mem0 vs verbatim (partial) | ≥4 systems per paper-topics checklist |
+| Threats / limitations | Section below | External review dedicated to OpMem only |
+
+**Publication verdict:** **not ready** for submission. Closest of the three paper tracks in [inventory-reproduction-plan-2026-09-18.md](../inventory-reproduction-plan-2026-09-18.md) (merge via PR #188), but requires manuscript, expanded systems, and frozen artifact bundle.
+
+---
+
+## Reproducible protocol
+
+### Scope
+
+OpMem measures **operational correctness** via deterministic operation scripts over `remember` / `recall` / `revise` / `forget`. Scoring is **binary pass/fail** per task on `GET /memories/search` results (Brainy adapter), not product `POST /recall` and not LLM-judge accuracy.
+
+### Task set (v0 + `upd03`)
+
+| Category | Task IDs | Count |
+| --- | --- | ---: |
+| suppression | `sup01_basic_forget`, `sup02_targeted_forget`, `sup03_durable_forget` | 3 |
+| correction | `cor01_basic_revision`, `cor02_correction_stickiness`, `cor03_revised_retrievable` | 3 |
+| isolation | `iso01_subject_isolation`, `iso02_tenant_isolation`, `iso03_forget_isolated` | 3 |
+| staleness | `upd01_stale_fact`, `upd02_preference_change`, `upd03_state_supersession` | 3 |
+| idempotency | `dup01_idempotent_remember` | 1 |
+| **Total** | | **13** |
+
+`upd03_state_supersession.json` added after the July 12-task public draft; it is the sole cardinality change for the 2026-08-15 pin.
+
+### Systems and mappings
+
+| Adapter | `remember` | `recall` | `revise` | `forget` |
+| --- | --- | --- | --- | --- |
+| Brainy | `POST /ingest` | `GET /memories/search` | `POST /memories/{id}/correct` | `POST /memories/{id}/suppress` |
+| Verbatim | in-process append | token overlap rank | in-place content replace | delete by id |
+| Mem0 Platform | `POST /v1/memories/` | `POST /v2/memories/search/` | `PUT /v1/memories/{id}/` | `DELETE /v1/memories/{id}/` |
+
+Hermeticity: Brainy and Mem0 adapters prefix tenants per task (`opmem-{nonce}-{task}`) so shared backends do not collide.
+
+### Assertion semantics
+
+Case-insensitive **substring** match on ranked search `content` fields. See [opmem-spec.md](../opmem-spec.md) for `min_results`, `max_results`, `top_contains`, `must_exclude`, etc.
+
+### Runner exit code
+
+`evals/run_opmem.py` exits **non-zero only on infrastructure errors** (HTTP/adapter exceptions), not on task `fail`. CI `TestOpMemBenchmarkAgainstHTTPServer` therefore proves **harness health**, not that Brainy passes 13/13, unless the test is extended to parse JSON (it does not today).
+
+### Commands (no external LLM required for Brainy+verbatim)
+
+```bash
+# CI-equivalent (embedded Postgres + in-process API; no Mem0 key)
+go test ./internal/api/ -run TestOpMemBenchmarkAgainstHTTPServer -count=1
+
+# Local API (start API+worker per CONTRIBUTING.md / AGENTS.md)
+unset BRAINY_API_KEYS BRAINY_REQUIRE_API_KEY
+export BRAINY_ENV=local
+python3 evals/run_opmem.py --systems verbatim,brainy --base-url http://127.0.0.1:8080 \
+  --json-out /tmp/opmem-report.json
+
+# Mem0 Platform counter-run (requires MEM0_API_KEY; costs API calls)
+MEM0_API_KEY=... python3 evals/run_opmem.py --systems mem0 \
+  --json-out /tmp/opmem-mem0-report.json
+```
+
+### Pin fields to record on any published table
+
+| Field | 2026-08-15 freeze value |
+| --- | --- |
+| Brainy git SHA | `1b5ab3e` (not re-validated on `bbe55f7` in this inventory) |
+| Fixture set | 13 files under `fixtures/opmem/` (git tree at SHA) |
+| Mem0 surface | **Platform** API via `Mem0OpAdapter` — **not** Mem0 OSS |
+| Mem0 run date | 2026-08-15 |
+| Brainy env (fresh local pin) | dedicated DB `brainy_bench`, `BRAINY_RECALL_LLM=1`, `BRAINY_USE_RECALL=1`, worker concurrency 8 |
+| Scoring endpoint | `/memories/search` (not `/recall`) |
+
+Artifacts: [opmem-fresh-local-20260815.md](../../benchmarks/artifacts/opmem-fresh-local-20260815.md), [opmem-mem0-fresh-20260815.md](../../benchmarks/artifacts/opmem-mem0-fresh-20260815.md).
+
+---
+
+## Exact claims and evidence table
+
+| Claim | Allowed wording | Evidence in repo | Gaps / caveats |
+| --- | --- | --- | --- |
+| Brainy passes all **13** OpMem tasks on search path | “13/13 on OpMem v0+`upd03` at SHA `1b5ab3e`” | [opmem-fresh-local-20260815.md](../../benchmarks/artifacts/opmem-fresh-local-20260815.md); [opmem-integrity-20260819.json](../../benchmarks/artifacts/opmem-integrity-20260819.json) (per-task `passed: true`) | Not re-run on `bbe55f7`; no committed Brainy+Mem0 combined JSON |
+| Mem0 Platform **10/13** on same fixtures | “10/13 Platform, 2026-08-15; fails `cor02`, `sup03`, `upd02`” | [opmem-mem0-fresh-20260815.md](../../benchmarks/artifacts/opmem-mem0-fresh-20260815.md) | No per-task Mem0 JSON in git; not OSS; API may drift |
+| Brainy leads Mem0 on ops | Same-pin **only** with above pins | Category table in mem0-fresh artifact | Do not imply LoCoMo lead; do not use July **9/12** as current Mem0 |
+| Verbatim baseline **10/13** | From Brainy fresh run same day | [opmem-fresh-local-20260815.md](../../benchmarks/artifacts/opmem-fresh-local-20260815.md) | Not Mem0-comparable headline |
+| July **12/12 vs 9/12** | Historical staging only | [posts/2026-07-opmem-v0.md](../posts/2026-07-opmem-v0.md), [opmem-staging-vs-mem0.json](../../benchmarks/opmem-staging-vs-mem0.json) | **Stale as current** — 12-task set, pre-`upd03` |
+| CI guarantees Brainy 13/13 | **Do not claim** | `TestOpMemBenchmarkAgainstHTTPServer` | Exit code ignores task failures |
+| “Beats Mem0” on conversational recall | **Forbidden** without same-pin LoCoMo | README / cycle-closeout | OpMem lead ≠ LoCoMo lead |
+
+---
+
+## Independent evidence review (13/13 vs 10/13)
+
+Performed from static repo artifacts only (no re-run).
+
+### Fixture cardinality
+
+`fixtures/opmem/*.json` → **13** files. Sorted runner order matches integrity JSON task list (13 entries).
+
+### Brainy 13/13
+
+- [opmem-fresh-local-20260815.md](../../benchmarks/artifacts/opmem-fresh-local-20260815.md): aggregate **13/13**, category splits correction/isolation/suppression **3/3**, staleness **3/3**, idempotency **1/1**.
+- [opmem-integrity-20260819.json](../../benchmarks/artifacts/opmem-integrity-20260819.json): 13 results, each `brainy.passed: true`, `failures: []`.
+
+**Not inflated:** we do not extrapolate to `bbe55f7` without a new run; integrity JSON is Brainy-only, not the full three-system counter-run.
+
+### Mem0 10/13
+
+- [opmem-mem0-fresh-20260815.md](../../benchmarks/artifacts/opmem-mem0-fresh-20260815.md): **10/13 (76.9%)**; failures documented as `cor02`, `sup03`, `upd02` with mechanism notes (ruby stickiness, forget leak, sms preference).
+- Arithmetic: 13 − 3 failures = 10 passes. Category table in artifact sums to 10 passed tasks.
+
+**Not inflated:** Platform-only; no reproducible OSS pin; no machine-readable Mem0 per-task log in git to independently re-score without re-running.
+
+### Verbatim cross-check (same Brainy run doc)
+
+Brainy fresh doc reports verbatim **10/13** — shows suite discriminates; does not change Mem0 pin.
+
+### Older 12-task JSON
+
+[opmem-staging-vs-mem0.json](../../benchmarks/opmem-staging-vs-mem0.json) lists **12** tasks (staleness **2/2**, no `upd03`). Brainy **12/12**, Mem0 **9/12**. This is consistent with the July set; it must not be merged into the 13-task pin.
+
+---
+
+## Missing experiments and ablations
+
+From [opmem-spec.md](../opmem-spec.md) roadmap and [inventory-reproduction-plan-2026-09-18.md](../inventory-reproduction-plan-2026-09-18.md):
+
+| Experiment | Why it matters | Repo state |
+| --- | --- | --- |
+| Mem0 **OSS** same 13 tasks | Fair open comparison | Not run; Platform 10/13 explicit |
+| Zep / Letta / LangMem adapters | Paper-topics ≥4 systems | Not implemented |
+| v1 task expansion (~30) | expiry, async visibility, races | Not built |
+| `revise` vs `POST .../supersede` | ENG-86 supersession API exists | OpMem still maps revise → `/correct` |
+| Brainy `/recall` vs `/search` on OpMem | Product path may differ | Harness fixes search lane only |
+| Ablation: correction vs suppression vs rank-only | Explain mechanism | No harness |
+| Re-run on current `dev` SHA | Current merge bar | **Done** — `20533f2` pin (`opmem-pin-20260918.json`) |
+| Commit frozen 3-system JSON (per-task) | Artifact review | **Done** — `docs/benchmarks/artifacts/opmem-pin-20260918.json` |
+| Statistical stability | 13 tasks, single run | **Partial** — 5× repeat in `opmem-stability.json`; only infra-error=0 runs count; 9/10 with errors = invalid/incomplete (not a second valid score) |
+
+---
+
+## Threats to validity
+
+1. **Small n:** 13 binary tasks; wide confidence intervals; one-task swings change percentages (~7.7% per task).
+2. **Substring scoring:** Passing does not prove semantic correctness or safe production answers.
+3. **Search vs recall:** Pins use `/memories/search`; product agents may use `/recall` with different ranking/abstain behavior.
+4. **CI gap:** Harness success does not assert pass rate (see protocol).
+5. **Mem0 Platform drift:** Vendor API/version changes can move 10/13 without Brainy changing.
+6. **Mem0 ≠ OSS:** Headline competitor number may not reproduce for readers on open-source Mem0.
+7. **English, domain-neutral fixtures:** May not represent multilingual or vertical-pack content.
+8. **No concurrency / TTL:** v0 scripts are sequential; race and expiry bugs undetected.
+9. **Adapter fidelity:** Mapping neutral ops to vendor endpoints may not match recommended vendor patterns.
+10. **SHA staleness:** Public README cites 13/13 on `1b5ab3e` cycle; `bbe55f7` merge bar not re-measured for OpMem in inventory.
+11. **Authoring bias:** Fixtures authored by Brainy team; independent fixture audit not documented.
+12. **Semantic contracts:** `sup03` durable forget and staleness without explicit supersede are **normative choices**; failing systems may reflect different valid policies.
+
+---
+
+## Artifact instructions (reviewers and authors)
+
+### Minimal reproduction (Brainy + verbatim)
+
+1. Checkout git SHA to pin (document in report; inventory used `1b5ab3e` for last full pin).
+2. `go test ./internal/api/ -run TestOpMemBenchmarkAgainstHTTPServer -count=1`
+3. Optional: write JSON via `python3 evals/run_opmem.py --systems verbatim,brainy --base-url <api> --json-out opmem-report.json`
+
+### Full competitive table (authors)
+
+1. Run Brainy and Mem0 in separate invocations or one `--systems brainy,verbatim,mem0` with live API + `MEM0_API_KEY`.
+2. Commit **one** JSON with `benchmark`, `systems`, `results[]` per task, `summary`, and manifest: git SHA, date, Mem0 surface (Platform/OSS), env vars affecting retrieval.
+3. Update [docs/benchmarks/artifacts/](../../benchmarks/artifacts/) markdown summary; do not replace LoCoMo pins.
+
+### What not to ship
+
+- July **12/12 vs 9/12** as current (use 13-task pin or label historical).
+- Platform Mem0 as “Mem0” without qualifier.
+- OpMem lead as LoCoMo or LME lead.
+
+### Doc consistency check (maintainers)
+
+```bash
+# Spec and ladder should not say "12 tasks" for current v0
+rg -n 'v0: 12 tasks|12 tasks\)' docs/research/opmem-spec.md docs/research/public-bench-ladder.md
+
+# Fixture count should match spec
+find fixtures/opmem -name '*.json' | wc -l
+```
+
+---
+
+## Manuscript outline
+
+**Working title:** *Beyond recall: operational correctness for agent memory systems*
+
+1. **Abstract** — Retrieval benchmarks miss forget/correct/isolate/supersede; OpMem binary suite; Brainy 13/13 vs Mem0 Platform 10/13 on frozen 2026-08-15 pin (qualified).
+2. **Introduction** — Production failure modes; gap vs LoCoMo/LME/HaluMem.
+3. **Related work** — Retrieval QA benchmarks; memory product surveys; Mem0 staleness/update gaps (cite their reports, not as same-pin scores).
+4. **Neutral operation model** — remember/recall/revise/forget; actors; adapter contract.
+5. **Task suite** — 13 tasks, five categories; semantic contracts (`sup03`, staleness).
+6. **Methodology** — Substring assertions; hermetic tenants; Platform vs OSS boundary; search endpoint.
+7. **Results** — Table: Brainy, Mem0 Platform, verbatim; per-task failure taxonomy for Mem0.
+8. **Discussion** — What failures imply; policy disagreements on durable forget.
+9. **Limitations** — n=13, English, no concurrency, CI vs full pin, SHA age.
+10. **Artifact appendix** — Fixture URLs, runner commands, JSON schema.
+11. **Future work** — v1 tasks, more adapters, supersede mapping, `/recall` lane.
+
+Target venue: workshop or arXiv benchmark track per [paper-topics.md](../paper-topics.md).
+
+---
+
+## Overlap with draft PRs #185–#187
+
+| PR | Topic | OpMem overlap | Consolidation |
+| --- | --- | --- | --- |
+| [#185](https://github.com/tryvinci/brainy/pull/185) | n=1540 LoCoMo handoff | Mentions keeping **13/13** green; no OpMem code | No merge conflict; publication package is independent |
+| [#186](https://github.com/tryvinci/brainy/pull/186) | eval-protocol-v2 staged LoCoMo | OpMem unchanged; LoCoMo qualification only | Do not conflate protocol-v2 with OpMem runner |
+| [#187](https://github.com/tryvinci/brainy/pull/187) | RRF retrieval flag | Test plan: OpMem **13/13** with `BRAINY_RETRIEVAL_RRF=0` before merge | Product change could affect search ranking; **re-run OpMem** after merge if RRF path touches shared rank |
+
+Related: PR [#188](https://github.com/tryvinci/brainy/pull/188) (`inventory-reproduction-plan-2026-09-18.md`) — §1.1 OpMem; this package **implements** Paper 1 deliverables from that inventory without duplicating the full three-paper inventory.
+
+---
+
+## Doc checks (this PR)
+
+- [x] `fixtures/opmem/` count = **13**
+- [x] [opmem-spec.md](../opmem-spec.md) updated to **13 tasks** and `upd03`
+- [x] [public-bench-ladder.md](../public-bench-ladder.md) L0 OpMem row → **13/13 vs 10/13** with artifact pointer
+- [x] [mem0-parity-gaps.md](../mem0-parity-gaps.md) executive OpMem row updated
+- [x] [master-plan.md](../master-plan.md) §1.1 OpMem row updated (13 tasks, current pin)
+- [x] [posts/2026-07-opmem-v0.md](../posts/2026-07-opmem-v0.md) supersession banner (July numbers preserved)
+- [ ] Historical reports (`opmem-baseline-report.md`, `launch-narrative.md`, staging JSON) — intentionally **dated**; not rewritten
+- [x] Re-run OpMem on current SHA — `opmem-manuscript-20260918.md`
+- [x] LaTeX manuscript — `docs/research/opmem/paper/`
+- [x] Phase 1 external lanes — adapters + dry-run worksheet ([external-lanes-setup.md](./external-lanes-setup.md), `cost-worksheet-dry-run.json`)
+- [ ] Mem0 OSS / Zep / Letta / LangMem — **live runs blocked** until credentials + measured estimate under ceiling (see worksheet)
+- [x] Mem0 Platform — **completed** on current pin (negligible API volume; `opmem-pin-20260918.json`)
diff --git a/docs/research/opmem/README.md b/docs/research/opmem/README.md
new file mode 100644
index 00000000..110d9e62
--- /dev/null
+++ b/docs/research/opmem/README.md
@@ -0,0 +1,16 @@
+# OpMem (operational memory correctness)
+
+Paper-track benchmark for whether a memory system **behaves correctly** under forget, correct, isolate, and supersede — not whether it recalls facts on LoCoMo.
+
+| Doc | Purpose |
+| --- | --- |
+| [opmem-spec.md](../opmem-spec.md) | Normative task schema, operations, scoring |
+| [PUBLICATION_READINESS.md](./PUBLICATION_READINESS.md) | **Publication package** (protocol, claims, gaps, threats, artifacts, outline) |
+| [paper/main.pdf](./paper/main.pdf) | Formatted manuscript (build: `make -C docs/research/opmem/paper pdf`) |
+| [external-lanes-setup.md](./external-lanes-setup.md) | Phase 1 external lane adapters (dry-run / cost worksheet) |
+| [posts/2026-07-opmem-v0.md](../posts/2026-07-opmem-v0.md) | July 2026 public draft (**12-task** numbers; superseded for pins) |
+| [inventory-reproduction-plan-2026-09-18.md](../inventory-reproduction-plan-2026-09-18.md) | Cross-paper inventory — **PR [#188](https://github.com/tryvinci/brainy/pull/188)** until merged to `dev` |
+
+**Code:** `fixtures/opmem/` (13 tasks) · `evals/run_opmem.py` · `evals/opmem_adapters.py` · CI `TestOpMemBenchmarkAgainstHTTPServer`
+
+**Current empirical pin (not re-run on `bbe55f7`):** Brainy **13/13** vs Mem0 Platform **10/13** (2026-08-15, SHA `1b5ab3e`). See [PUBLICATION_READINESS.md](./PUBLICATION_READINESS.md#independent-evidence-review-132-vs-1013).
diff --git a/docs/research/opmem/external-lanes-setup.md b/docs/research/opmem/external-lanes-setup.md
new file mode 100644
index 00000000..3da51b1c
--- /dev/null
+++ b/docs/research/opmem/external-lanes-setup.md
@@ -0,0 +1,84 @@
+# OpMem external lanes — setup
+
+Phase 1 ships adapters and dry-run operation counts. **Spend approval** uses the payload budget worksheet (`cost-worksheet-payload-budget.json`), not the legacy mock token meter in `cost-worksheet-dry-run.json`. Phase 2 live runs: `evals/run_opmem_external_live.py`.
+
+## Spend authority (USD)
+
+| Lane | Ceiling | Notes |
+| --- | ---: | --- |
+| `mem0-oss` | $1 | Local Qdrant + OpenAI for extract/embed |
+| `letta` | $1 | Self-hosted App Server + OpenAI |
+| `langmem` | $1 | Postgres/pgvector + OpenAI |
+| `zep` | **$0 paid** | Zep Cloud free tier only; stop if credits exhausted |
+| **Total** | **$3** | Sum of OpenAI-metered lanes |
+
+Rates used in worksheets (OpenAI list, per 1M tokens): `gpt-4.1-mini-2025-04-14` input $0.40 / output $1.60; `text-embedding-3-small` $0.02.
+
+## Pinned models
+
+- Chat: `gpt-4.1-mini-2025-04-14`
+- Embed: `text-embedding-3-small`
+
+If a lane cannot accept these pins, record the product constraint in the worksheet and **do not** substitute silently (Zep is documented as Zep-native only).
+
+## Environment variables (never commit values)
+
+| Lane | Variables |
+| --- | --- |
+| **mem0-oss** | `OPENAI_API_KEY` (vault); `QDRANT_URL` (default `http://127.0.0.1:6333`); optional `MEM0_OSS_CONFIG_PATH` |
+| **zep** | `ZEP_API_KEY`; optional `ZEP_API_URL` (default `https://api.getzep.com`) |
+| **letta** | `OPENAI_API_KEY`; `LETTA_BASE_URL` (default `http://127.0.0.1:8283`); optional `LETTA_APP_SERVER_TOKEN` (required for non-local servers) |
+| **langmem** | `OPENAI_API_KEY`; `DATABASE_URL` (Postgres + pgvector); **do not** set `LANGMEM_STORE=InMemoryStore` for measured runs |
+
+**Not used for mem0-oss lane:** `MEM0_API_KEY` (Platform).
+
+Deterministic namespaces: set `OPMEM_RUN_ID` (e.g. `opmem-pin-20260919`) for reproducible IDs.
+
+## Pinned Python packages
+
+Install once: `pip install -r evals/opmem_external/requirements-external-lanes.txt`
+
+| Component | Pin |
+| --- | --- |
+| Letta server | `letta==0.11.7` + `sqlite-vec` |
+| Letta client | `letta-client==0.1.307` |
+| Zep SDK | `zep-cloud==3.28.0` |
+
+Start local Letta (no Docker): `scripts/opmem-start-letta-server.sh`
+
+## Docker images (optional alternative)
+
+| Service | Image | Port |
+| --- | --- | ---: |
+| Qdrant (mem0-oss) | `qdrant/qdrant:v1.12.5` | 6333 |
+| Letta | prefer pip pin above; or `letta/letta` digest at live time | 8283 |
+| Postgres (langmem) | `pgvector/pgvector:pg17` | 5432 |
+
+## Phase 1 commands (no paid vendor API traffic)
+
+```bash
+# Dry-run all four lanes over 13 fixtures + cost worksheet
+python3 evals/run_opmem_cost_dry_run.py \
+  --out-dir docs/benchmarks/artifacts/opmem-external-lanes
+
+# Contract tests (mocked HTTP only)
+python3 -m unittest evals/opmem_external/test_lane_contract.py -v
+```
+
+Live harness:
+
+```bash
+pip install -r evals/opmem_external/requirements-external-lanes.txt
+scripts/opmem-start-letta-server.sh   # letta lane
+export DATABASE_URL=postgresql://opmem:opmem@127.0.0.1:5432/opmem_langmem
+export ZEP_API_KEY=...               # free tier only
+python3 evals/run_opmem_external_live.py --smoke --lanes mem0-oss,zep,letta,langmem
+python3 evals/run_opmem_external_live.py --full --lanes mem0-oss,zep,letta,langmem
+```
+
+## Fairness rules
+
+- Identical 13 tasks, step order, and substring scoring as `evals/run_opmem.py`.
+- Memory mutations only in adapters; no answer-generation calls in OpMem scoring.
+- `begin_task` resets lane namespace every task.
+- Infrastructure errors remain **invalid/incomplete** runs (not comparable task scores).
diff --git a/docs/research/opmem/paper/.gitignore b/docs/research/opmem/paper/.gitignore
new file mode 100644
index 00000000..c1ed4d59
--- /dev/null
+++ b/docs/research/opmem/paper/.gitignore
@@ -0,0 +1,7 @@
+*.aux
+*.bbl
+*.blg
+*.fdb_latexmk
+*.fls
+*.log
+*.out
diff --git a/docs/research/opmem/paper/Makefile b/docs/research/opmem/paper/Makefile
new file mode 100644
index 00000000..889f3285
--- /dev/null
+++ b/docs/research/opmem/paper/Makefile
@@ -0,0 +1,14 @@
+# Reproducible PDF build for the OpMem manuscript.
+.PHONY: pdf clean check
+
+MAIN = main
+LATEXMK = latexmk
+
+pdf:
+	$(LATEXMK) -pdf -interaction=nonstopmode -halt-on-error $(MAIN).tex
+
+check: pdf
+	@test -f $(MAIN).pdf
+
+clean:
+	$(LATEXMK) -C $(MAIN).tex
diff --git a/docs/research/opmem/paper/generated/README.md b/docs/research/opmem/paper/generated/README.md
new file mode 100644
index 00000000..8e3bef83
--- /dev/null
+++ b/docs/research/opmem/paper/generated/README.md
@@ -0,0 +1,5 @@
+# Generated manuscript fragments
+
+Most files here are produced by `evals/run_opmem_paper_experiments.py` / `scripts/run-opmem-manuscript.sh`.
+
+`external-lanes.tex` is **hand-maintained** (Phase~2 lane scores and Table~\ref{tab:external-lanes} caption). Caption wording: each lane runs the full 13-task script; Zep publication score is **3/9\*** on nine supported tasks (see `zep-score-interpretation-opmem-phase2-zep-full.json`). Long artifact filenames use `\artifact{...}` from `main.tex` (`xurl` + `\sloppy` block) so paths break inside the margin.
diff --git a/docs/research/opmem/paper/generated/combined-results.tex b/docs/research/opmem/paper/generated/combined-results.tex
new file mode 100644
index 00000000..fa4d8a02
--- /dev/null
+++ b/docs/research/opmem/paper/generated/combined-results.tex
@@ -0,0 +1,4 @@
+\evidence{completed}{
+Committed \texttt{opmem-primary.json} lists per-task outcomes for
+verbatim, brainy, mem0 with manifest fields (SHA, date, fixture count).
+}
diff --git a/docs/research/opmem/paper/generated/current-sha-results.tex b/docs/research/opmem/paper/generated/current-sha-results.tex
new file mode 100644
index 00000000..7557ade5
--- /dev/null
+++ b/docs/research/opmem/paper/generated/current-sha-results.tex
@@ -0,0 +1,5 @@
+\evidence{completed}{
+At git SHA \texttt{20533f2ca83e} (UTC 2026-09-18T23:12:29Z), embedded API harness:
+Brainy (search lane) \textbf{13/13}, verbatim \textbf{10/13}.
+Mem0 Platform (same run): \textbf{10/13} (API surface: Platform via \texttt{Mem0OpAdapter}).
+}
diff --git a/docs/research/opmem/paper/generated/external-lanes.tex b/docs/research/opmem/paper/generated/external-lanes.tex
new file mode 100644
index 00000000..9c51e905
--- /dev/null
+++ b/docs/research/opmem/paper/generated/external-lanes.tex
@@ -0,0 +1,26 @@
+\begin{table}[ht]
+\centering
+\small
+\begin{tabular}{@{}lcc@{}}
+\toprule
+Lane & OpMem score & Run \\
+\midrule
+Brainy (primary pin) & 13/13 & \texttt{opmem-pin-20260918} \\
+Mem0 Platform & 10/13 & same pin \\
+Mem0 OSS & 9/13 & \texttt{opmem-phase2-full-mem0} \\
+LangMem & 8/13 & \texttt{opmem-phase2-full1} \\
+Letta & 5/13 & \texttt{opmem-phase2-letta-full} \\
+Zep & 3/9$^{*}$ & \texttt{opmem-phase2-zep-full} \\
+\bottomrule
+\end{tabular}
+\caption{Phase~2 external lanes: each lane ran the full 13-task OpMem script; Zep is scored on \textbf{9 supported tasks}$^{*}$. $^{*}$Four forget fixtures (\texttt{iso03\_forget\_isolated}, \texttt{sup01--sup03}) are \emph{not supported} on the Zep lane; raw harness tally \textbf{3/13} unchanged on \artifact{live-full-opmem-phase2-zep-full.json}. Forget is not supported (\texttt{zep-cloud} 3.28 thread/graph API has no delete path the adapter can use).}
+\label{tab:external-lanes}
+\end{table}
+
+{\sloppy
+\evidence{completed}{Mem0 \textbf{OSS} (local Qdrant path, \texttt{mem0ai} 2.1.0): \textbf{9/13} on run \texttt{opmem-phase2-full-mem0} (0 infra errors). Artifact: \artifact{live-full-opmem-phase2-full-mem0.json}.}
+\evidence{completed}{LangMem (Postgres 16 + \texttt{pgvector} 0.6): \textbf{8/13} on run \texttt{opmem-phase2-full1} (0 infra errors). Artifact: \artifact{live-full-opmem-phase2-full1.json}.}
+\evidence{completed}{Letta (\texttt{letta==0.11.7}): \textbf{5/13} on run \texttt{opmem-phase2-letta-full} (0 infra errors). Artifact: \artifact{live-full-opmem-phase2-letta-full.json}.}
+\evidence{completed}{Zep (\$0 paid, \texttt{zep-cloud} 3.28.0): \textbf{3/9$^{*}$} on supported tasks from run \texttt{opmem-phase2-zep-full} (0 infra errors). \emph{Not supported} (not scored as failures): \artifact{iso03\_forget\_isolated}, \artifact{sup01\_basic\_forget}, \artifact{sup02\_targeted\_forget}, \artifact{sup03\_durable\_forget}. Raw harness: \textbf{3/13}.\newline Artifact: \artifact{live-full-opmem-phase2-zep-full.json}; interpretation: \artifact{zep-score-interpretation-opmem-phase2-zep-full.json}.}
+\evidence{completed}{Payload budget worksheet: worst-case OpenAI \textbf{\$0.15} vs \$3 ceiling (\artifact{cost-worksheet-payload-budget.json}).}
+\par}
diff --git a/docs/research/opmem/paper/generated/lane-ablation.tex b/docs/research/opmem/paper/generated/lane-ablation.tex
new file mode 100644
index 00000000..917e309e
--- /dev/null
+++ b/docs/research/opmem/paper/generated/lane-ablation.tex
@@ -0,0 +1,17 @@
+\begin{table}[h]
+\centering
+\caption{Brainy recall lane and revise mapping (SHA 20533f2ca83e).}
+\begin{tabular}{lc}
+\toprule
+Configuration & Overall \\
+\midrule
+brainy & 13/13 \\
+brainy-recall & 9/13 \\
+brainy-supersede & 11/13 \\
+\bottomrule
+\end{tabular}
+\end{table}
+Search (\texttt{/memories/search}) is the normative pin lane.
+\texttt{brainy-recall} uses \texttt{POST /recall} (context mode).
+Correction tasks on the recall lane may return HTTP 409 during \texttt{/correct}
+(infrastructure errors in artifact JSON).
diff --git a/docs/research/opmem/paper/generated/mechanism-ablation.tex b/docs/research/opmem/paper/generated/mechanism-ablation.tex
new file mode 100644
index 00000000..a373ff5a
--- /dev/null
+++ b/docs/research/opmem/paper/generated/mechanism-ablation.tex
@@ -0,0 +1,16 @@
+Category pass rates separate correction/suppression/staleness contracts;
+verbatim approximates rank-only recall without governed lifecycle ops.
+\begin{table}[h]
+\centering
+\begin{tabular}{lcc}
+\toprule
+Category & Brainy & Verbatim (rank-only) \\
+\midrule
+correction & 3/3 & 2/3 \\
+idempotency & 1/1 & 0/1 \\
+isolation & 3/3 & 3/3 \\
+staleness & 3/3 & 3/3 \\
+suppression & 3/3 & 2/3 \\
+\bottomrule
+\end{tabular}
+\end{table}
diff --git a/docs/research/opmem/paper/generated/stability.tex b/docs/research/opmem/paper/generated/stability.tex
new file mode 100644
index 00000000..b87808dc
--- /dev/null
+++ b/docs/research/opmem/paper/generated/stability.tex
@@ -0,0 +1,13 @@
+\evidence{completed}{
+Five consecutive Brainy search-lane repeats on one embedded server session.
+\textbf{Valid runs only} (infra errors $=0$): 2/5; scores: 13/13.
+}
+\begin{itemize}
+\item Repeat 1: 13/13 (infra errors=0)
+\item Repeat 2: 9/10 (infra errors=3)
+\item Repeat 3: 9/10 (infra errors=3)
+\item Repeat 4: 13/13 (infra errors=0)
+\item Repeat 5: 9/10 (infra errors=3)
+\end{itemize}
+Repeats with infra errors $>0$ (3 of 5) are \textbf{invalid/incomplete}: the harness omits errored tasks from the denominator, so aggregates such as 9/10 are \emph{not} comparable to a clean 13/13 pin.
+Use isolated \texttt{TestOpMemBenchmarkAgainstHTTPServer} for merge-gate stability.
diff --git a/docs/research/opmem/paper/generated/supersede-ablation.tex b/docs/research/opmem/paper/generated/supersede-ablation.tex
new file mode 100644
index 00000000..b9898a39
--- /dev/null
+++ b/docs/research/opmem/paper/generated/supersede-ablation.tex
@@ -0,0 +1,5 @@
+\evidence{completed}{
+Mapping \texttt{revise} $\rightarrow$ \texttt{POST .../supersede} yields
+\textbf{11/13} vs \textbf{13/13} with \texttt{/correct} on the search lane
+(same SHA). Failures are task-level assertion misses, not missing runs.
+}
diff --git a/docs/research/opmem/paper/main.pdf b/docs/research/opmem/paper/main.pdf
new file mode 100644
index 00000000..4b6da2ad
Binary files /dev/null and b/docs/research/opmem/paper/main.pdf differ
diff --git a/docs/research/opmem/paper/main.tex b/docs/research/opmem/paper/main.tex
new file mode 100644
index 00000000..dae0c7c0
--- /dev/null
+++ b/docs/research/opmem/paper/main.tex
@@ -0,0 +1,214 @@
+\documentclass[11pt]{article}
+\usepackage[margin=1in]{geometry}
+\usepackage{booktabs}
+\usepackage{hyperref}
+\usepackage{xurl}
+\urlstyle{tt}
+\newcommand{\artifact}[1]{\url{#1}}
+\usepackage{longtable}
+\usepackage{xcolor}
+\usepackage{listings}
+\usepackage{graphicx}
+
+\newcommand{\todo}[1]{\textcolor{red}{\textbf{TODO:} #1}}
+\newcommand{\evidence}[2]{\textbf{[#1]} #2}
+
+\title{Beyond Recall: Operational Correctness for Agent Memory Systems\\
+\large OpMem v0 Benchmark and Empirical Pins}
+\author{Brainy Research}
+\date{\today}
+
+\begin{document}
+\maketitle
+
+\begin{abstract}
+Retrieval benchmarks such as LoCoMo and LongMemEval measure whether an agent recalls
+the right fact, but not whether a memory \emph{system} honors forget, correct, isolate,
+and supersede semantics in production.
+We present \textbf{OpMem v0}, a 13-task, pass/fail suite over neutral
+\texttt{remember}/\texttt{recall}/\texttt{revise}/\texttt{forget} scripts with
+substring-based assertions on ranked results.
+On the frozen 2026-08-15 pin (git SHA \texttt{1b5ab3e}), Brainy passes \textbf{13/13}
+via \texttt{GET /memories/search} while Mem0 \textbf{Platform} passes \textbf{10/13}
+(failures: \texttt{cor02}, \texttt{sup03}, \texttt{upd02}).
+This manuscript reports repository evidence, a reproducible protocol, threats to validity,
+and new embedded-harness re-measurements on the current merge-pin git revision where completed.
+We do \emph{not} claim conversational-recall superiority over Mem0; OpMem operational
+leadership is orthogonal to LoCoMo J-scores.
+\end{abstract}
+
+\section{Introduction}
+Production agents fail when suppressed memories leak after re-ingestion, corrections do not
+stick, tenants bleed across isolation boundaries, or stale facts outrank updates.
+Conversational QA benchmarks~\cite{maharana2024locomo,longmemeval2024} do not encode these
+lifecycle contracts; hallucination-oriented memory benchmarks such as HaluMem~\cite{chen2025halumem}
+focus on factual faithfulness rather than governed delete/correct semantics.
+
+OpMem targets \textbf{operational correctness}: deterministic scripts, binary scoring, and
+explicit adapter mappings to each system's native API surface~\cite{brainyopmem2026}.
+Section~\ref{sec:limitations} lists threats; Section~\ref{sec:results} reports primary,
+historical, and Phase~2 external-lane pins (Table~\ref{tab:external-lanes}).
+
+\section{Related work}
+\paragraph{Retrieval QA.}
+LoCoMo~\cite{maharana2024locomo} and LongMemEval~\cite{longmemeval2024} score end-to-end
+answer quality over long conversations; they are not substitutes for lifecycle tests.
+
+\paragraph{Memory products.}
+Mem0 documents accumulation-oriented designs and names staleness as an open problem~\cite{mem0state2026}.
+We report Mem0 \textbf{Platform} numbers only on OpMem fixtures; we do not import Mem0 blog
+LoCoMo percentages as same-pin scores.
+
+\paragraph{External memory lanes (OpMem v0).}
+Mem0 \textbf{Platform} on OpMem v0 is \textbf{completed} at negligible API volume (13 tasks, 47 scripted steps; Section~\ref{sec:external}).
+Phase~2 counter-runs (after cost worksheet approval) are \textbf{completed} for Mem0 OSS, LangMem, Letta, and Zep; scores in Table~\ref{tab:external-lanes}: \textbf{9/13}, \textbf{8/13}, \textbf{5/13}, and \textbf{3/9$^{*}$} respectively (Zep raw harness \textbf{3/13}; \$0 paid).
+Paid or vendor-hosted lanes require a written estimate before approval; that gate was not a categorical denial of spend.
+
+\section{Neutral operation model}
+\label{sec:model}
+Actors are \texttt{(tenant, subject)} pairs.
+Tasks are operation scripts; \texttt{revise} and \texttt{forget} reference prior
+\texttt{remember} step ids.
+Adapters namespace tenants per task (\texttt{opmem-\{nonce\}-\{task\}}) for hermetic runs.
+
+\begin{table}[h]
+\centering
+\caption{Adapter mappings (normative; from \texttt{PUBLICATION\_READINESS.md}).}
+\begin{tabular}{llll}
+\toprule
+Op & Brainy & Verbatim baseline & Mem0 Platform \\
+\midrule
+remember & \texttt{POST /ingest} & append & \texttt{POST /v1/memories/} \\
+recall & \texttt{GET /memories/search} & token overlap & \texttt{POST /v2/memories/search/} \\
+revise & \texttt{POST .../correct} & in-place replace & \texttt{PUT /v1/memories/\{id\}/} \\
+forget & \texttt{POST .../suppress} & delete & \texttt{DELETE /v1/memories/\{id\}/} \\
+\bottomrule
+\end{tabular}
+\end{table}
+
+\section{Task suite}
+OpMem v0 comprises 13 JSON fixtures under \texttt{fixtures/opmem/}: three suppression,
+three correction, three isolation, three staleness (including \texttt{upd03\_state\_supersession}),
+and one idempotency task.
+Assertions use case-insensitive substring checks on ranked \texttt{content} fields
+(\texttt{min\_results}, \texttt{top\_contains}, \texttt{must\_exclude}, etc.).
+
+\section{Methodology}
+\subsection{Scoring lane}
+Primary pins use Brainy \texttt{/memories/search}, not product \texttt{POST /recall}
+(Section~\ref{sec:lane} reports a \texttt{/recall} lane ablation).
+The harness (\texttt{evals/run\_opmem.py}) exits non-zero only on infrastructure errors,
+not task failures---CI proves runner health, not 13/13, unless JSON is parsed.
+
+\subsection{Reproduction commands}
+\begin{lstlisting}[basicstyle=\ttfamily\small]
+go test ./internal/api/ -run TestOpMemBenchmarkAgainstHTTPServer -count=1
+python3 evals/run_opmem.py --systems verbatim,brainy \
+  --base-url http://127.0.0.1:8080 --json-out opmem-report.json
+\end{lstlisting}
+Paper experiments: \texttt{scripts/run-opmem-manuscript.sh} and
+\texttt{evals/run\_opmem\_paper\_experiments.py}.
+
+\section{Results}
+\label{sec:results}
+
+\subsection{Frozen 2026-08-15 pin (historical, SHA \texttt{1b5ab3e})}
+\evidence{completed}{
+Brainy \textbf{13/13}; verbatim \textbf{10/13}; Mem0 Platform \textbf{10/13}
+(artifacts \texttt{opmem-fresh-local-20260815.md}, \texttt{opmem-mem0-fresh-20260815.md}).
+}
+
+\begin{table}[h]
+\centering
+\caption{Category breakdown on 2026-08-15 Mem0 Platform counter-run.}
+\begin{tabular}{lrr}
+\toprule
+Category & Mem0 Platform & Brainy (same cycle) \\
+\midrule
+correction & 2/3 & 3/3 \\
+isolation & 3/3 & 3/3 \\
+suppression & 2/3 & 3/3 \\
+staleness & 2/3 & 3/3 \\
+idempotency & 1/1 & 1/1 \\
+\midrule
+\textbf{overall} & \textbf{10/13} & \textbf{13/13} \\
+\bottomrule
+\end{tabular}
+\end{table}
+
+Mem0 failures on that pin: \texttt{cor02} (correction stickiness),
+\texttt{sup03} (durable forget leak), \texttt{upd02} (preference ranking).
+
+\subsection{Current-SHA re-measurement}
+\label{sec:current}
+\input{generated/current-sha-results.tex}
+
+\subsection{Combined machine-readable artifact}
+\input{generated/combined-results.tex}
+
+\subsection{Recall vs search lane}
+\label{sec:lane}
+\input{generated/lane-ablation.tex}
+
+\subsection{Correct vs supersede mapping}
+\input{generated/supersede-ablation.tex}
+
+\subsection{Mechanism breakdown}
+\input{generated/mechanism-ablation.tex}
+
+\subsection{Statistical stability}
+\input{generated/stability.tex}
+
+\subsection{External systems}
+\label{sec:external}
+\evidence{completed}{Mem0 \textbf{Platform}: 10/13 on SHA \texttt{20533f2} (\texttt{opmem-pin-20260918.json}); same headline as 2026-08-15 pin. Completed under negligible Platform API use (full suite $\approx$47 scripted steps).}
+\input{generated/external-lanes.tex}
+
+\subsection{Cost-estimation plan (before any new paid run)}
+\label{sec:cost-plan}
+OpMem v0 has 13 tasks and 47 scripted steps (\texttt{fixtures/opmem/}). Before any new paid lane run, prepare billing and approval:
+\begin{itemize}
+\item \textbf{Phase 1--2} --- \texttt{evals/opmem\_external/}: deterministic namespaces, per-task reset, payload budget (\texttt{cost-worksheet-payload-budget.json}), live harness \texttt{run\_opmem\_external\_live.py}. Setup: \texttt{external-lanes-setup.md}.
+\item \textbf{Mem0 OSS} --- Local Qdrant + \texttt{OPENAI\_API\_KEY}; ceiling \$1; dry-run projected OpenAI spend in worksheet.
+\item \textbf{Zep} --- Zep Cloud free tier only (\$0 paid); cannot pin OpenAI models (product constraint recorded).
+\item \textbf{Letta} --- Self-hosted App Server; ceiling \$1; no legacy Docker API or Letta Cloud paid plan.
+\item \textbf{LangMem} --- Postgres/pgvector; ceiling \$1; no \texttt{InMemoryStore} on measured runs.
+\end{itemize}
+No dollar totals are asserted here until rate cards and deployment mode are pinned; approval requires the completed worksheet, not a categorical spend ban.
+
+\section{Discussion}
+OpMem failures on Mem0 align with accumulation-oriented update semantics:
+corrections and durable forgets are not first-class lifecycle events.
+Verbatim baseline passes 10/13, showing the suite discriminates rank-only stores from
+governed suppression/correction.
+Policy disagreements remain (\texttt{sup03} durable forget is a normative contract).
+
+\section{Limitations}
+\label{sec:limitations}
+\begin{enumerate}
+\item Small $n=13$; each task is $\sim$7.7\% of the aggregate score.
+\item Substring scoring does not guarantee semantic safety.
+\item Search vs \texttt{/recall} product path may diverge (lane ablation quantifies on Brainy).
+\item English, domain-neutral fixtures; no concurrency or TTL tests.
+\item Mem0 Platform API drift; Phase~2 OSS/LangMem/Letta/Zep pins are single-pass, small $n$. Zep forget fixtures are \emph{not supported} on the lane (Table~\ref{tab:external-lanes} footnote); publication line \textbf{3/9$^{*}$} vs raw harness \textbf{3/13}.
+\item Fixture authorship bias; no independent external audit documented.
+\end{enumerate}
+
+\section{Artifact appendix}
+Committed run manifests and JSON under
+\texttt{docs/benchmarks/artifacts/opmem-manuscript-*}:
+\texttt{run-manifest.json}, \texttt{opmem-primary.json},
+\texttt{opmem-lane-ablation.json}, \texttt{opmem-stability.json},
+\texttt{opmem-mechanism.json}, and \texttt{commands.log}.
+
+Build PDF: \texttt{make -C docs/research/opmem/paper pdf}.
+
+\section{Future work}
+v1 task expansion ($\sim$30 tasks), Zep forget/delete mapping if a supported API path appears,
+supersede-native revise mapping as default where lineage is required,
+and fail-closed CI assertion of 13/13 on merge pins.
+
+\bibliographystyle{plain}
+\bibliography{references}
+
+\end{document}
diff --git a/docs/research/opmem/paper/references.bib b/docs/research/opmem/paper/references.bib
new file mode 100644
index 00000000..8e50e0b5
--- /dev/null
+++ b/docs/research/opmem/paper/references.bib
@@ -0,0 +1,38 @@
+@article{maharana2024locomo,
+  title   = {{LoCoMo}: Evaluating Very Long-Term Conversational Memory of LLM Agents},
+  author  = {Maharana, Adyasha and others},
+  journal = {arXiv preprint arXiv:2402.17710},
+  year    = {2024},
+  url     = {https://arxiv.org/abs/2402.17710}
+}
+
+@article{longmemeval2024,
+  title   = {{LongMemEval}: Benchmarking Long-Term Memory in LLM Chat Assistants},
+  author  = {Wu, Howard Yen and others},
+  journal = {arXiv preprint arXiv:2410.10819},
+  year    = {2024},
+  url     = {https://arxiv.org/abs/2410.10819}
+}
+
+@article{chen2025halumem,
+  title   = {{HaluMem}: Evaluating Hallucinations in Memory Systems of Agents},
+  author  = {Chen, Ding and Niu, Simin and Li, Kehang and Liu, Peng and Zheng, Xiangping and Tang, Bo and Li, Xinchi and Xiong, Feiyu and Li, Zhiyu},
+  journal = {arXiv preprint arXiv:2511.03506},
+  year    = {2025},
+  url     = {https://arxiv.org/abs/2511.03506}
+}
+
+@misc{mem0state2026,
+  title        = {State of {AI} Agent Memory (2026)},
+  author       = {{Mem0}},
+  howpublished = {\url{https://mem0.ai/blog/state-of-ai-agent-memory-2026}},
+  year         = {2026},
+  note         = {Vendor context on staleness and cross-session evolution; not an OpMem same-pin score.}
+}
+
+@misc{brainyopmem2026,
+  title        = {{OpMem} v0 specification and harness},
+  author       = {{Brainy}},
+  howpublished = {\url{https://github.com/tryvinci/brainy} (\texttt{docs/research/opmem-spec.md})},
+  year         = {2026}
+}
diff --git a/docs/research/posts/2026-07-opmem-v0.md b/docs/research/posts/2026-07-opmem-v0.md
index 1840a174..685e5da6 100644
--- a/docs/research/posts/2026-07-opmem-v0.md
+++ b/docs/research/posts/2026-07-opmem-v0.md
@@ -1,5 +1,7 @@
 # OpMem v0: operational memory correctness beyond LOCOMO recall
 
+> **Superseded for pins:** This draft reports the **12-task** staging set (pre-`upd03`). Current empirical pin: Brainy **13/13** vs Mem0 Platform **10/13** on 13 fixtures ([opmem-fresh-local-20260815](../../benchmarks/artifacts/opmem-fresh-local-20260815.md)). See [opmem/PUBLICATION_READINESS.md](../opmem/PUBLICATION_READINESS.md).
+
 **Status:** Public research draft (Paper 1 track)  
 **Date:** 2026-07-29  
 **Systems:** Brainy (staging) · Mem0 Platform · Verbatim baseline  
diff --git a/docs/research/public-bench-ladder.md b/docs/research/public-bench-ladder.md
index b8a38686..7efeab6c 100644
--- a/docs/research/public-bench-ladder.md
+++ b/docs/research/public-bench-ladder.md
@@ -49,7 +49,7 @@ L6  Vertical “MarketingMem” public track (Brainy’s differentiation)
 | Metric | Brainy | Mem0 | Artifact |
 | --- | ---: | ---: | --- |
 | Parity | 4/4 | 4/4 | [staging report](../benchmarks/staging-competitive-report.md) |
-| OpMem | 12/12 | 9/12 | same |
+| OpMem | **13/13** (`1b5ab3e` pin) | **10/13** Platform (`2026-08-15`) | [opmem-fresh-local-20260815](../benchmarks/artifacts/opmem-fresh-local-20260815.md) · July staging **12/12 vs 9/12** is pre-`upd03` |
 | Marketing vertical | 16/16 | N/A | [moat](../benchmarks/marketing-moat-report.md) |
 
 ---
diff --git a/evals/opmem_adapters.py b/evals/opmem_adapters.py
index b7d76d78..055552ad 100644
--- a/evals/opmem_adapters.py
+++ b/evals/opmem_adapters.py
@@ -62,9 +62,23 @@ class BrainyAdapter:
 
     name = "brainy"
 
-    def __init__(self, base_url: str) -> None:
+    def __init__(
+        self,
+        base_url: str,
+        *,
+        recall_lane: str = "search",
+        revise_mode: str = "correct",
+    ) -> None:
         self.base_url = base_url.rstrip("/")
         self._run_nonce = str(int(time.time()))
+        self._recall_lane = recall_lane
+        self._revise_mode = revise_mode
+        if recall_lane == "recall":
+            self.name = "brainy-recall"
+        elif revise_mode == "supersede":
+            self.name = "brainy-supersede"
+        else:
+            self.name = "brainy"
 
     def available(self) -> bool:
         return True
@@ -88,6 +102,24 @@ def remember(self, actor: tuple[str, str], content: str) -> list[str]:
 
     def recall(self, actor: tuple[str, str], query: str) -> list[dict]:
         tenant, subject = self._scope(actor)
+        if self._recall_lane == "recall":
+            body = post_json(self.base_url, "/recall", {
+                "tenant_id": tenant,
+                "subject_id": subject,
+                "q": query,
+                "mode": "context",
+                "top_k": 20,
+            })
+            results = [
+                {"id": r["memory_id"], "content": r["content"]}
+                for r in body.get("memories", [])
+                if isinstance(r, dict) and r.get("content")
+            ]
+            if not results:
+                block = (body.get("context_block") or body.get("answer") or "").strip()
+                if block:
+                    results = [{"id": "recall-context", "content": block}]
+            return results
         body = get_json(self.base_url, "/memories/search", {
             "tenant_id": tenant,
             "subject_id": subject,
@@ -104,10 +136,18 @@ def forget(self, actor: tuple[str, str], memory_ids: list[str]) -> None:
     def revise(self, actor: tuple[str, str], memory_ids: list[str], content: str) -> None:
         tenant, subject = self._scope(actor)
         query = urllib.parse.urlencode({"tenant_id": tenant, "subject_id": subject})
+        payload = {"content": content, "source_text": content}
+        if self._revise_mode == "supersede":
+            post_json(
+                self.base_url,
+                f"/memories/{memory_ids[0]}/supersede?{query}",
+                payload,
+            )
+            return
         post_json(
             self.base_url,
             f"/memories/{memory_ids[0]}/correct?{query}",
-            {"content": content, "source_text": content},
+            payload,
         )
 
 
diff --git a/evals/opmem_external/__init__.py b/evals/opmem_external/__init__.py
new file mode 100644
index 00000000..ba602d08
--- /dev/null
+++ b/evals/opmem_external/__init__.py
@@ -0,0 +1,21 @@
+"""OpMem external memory lane adapters (Phase 1: dry-run + instrumentation)."""
+
+from opmem_external.langmem_store import LangMemLaneAdapter
+from opmem_external.letta_appserver import LettaAppServerLaneAdapter
+from opmem_external.mem0_oss import Mem0OSSLaneAdapter
+from opmem_external.zep_cloud import ZepCloudLaneAdapter
+
+LANE_ADAPTERS = {
+    "mem0-oss": Mem0OSSLaneAdapter,
+    "zep": ZepCloudLaneAdapter,
+    "letta": LettaAppServerLaneAdapter,
+    "langmem": LangMemLaneAdapter,
+}
+
+__all__ = [
+    "LANE_ADAPTERS",
+    "LangMemLaneAdapter",
+    "LettaAppServerLaneAdapter",
+    "Mem0OSSLaneAdapter",
+    "ZepCloudLaneAdapter",
+]
diff --git a/evals/opmem_external/cost_budget.py b/evals/opmem_external/cost_budget.py
new file mode 100644
index 00000000..1e6e133a
--- /dev/null
+++ b/evals/opmem_external/cost_budget.py
@@ -0,0 +1,302 @@
+"""Conservative pre-call token budget from serialized request payloads (fixtures × lane)."""
+
+from __future__ import annotations
+
+import json
+from dataclasses import dataclass, field
+from pathlib import Path
+from typing import Any, Iterable
+
+import tiktoken
+
+from opmem_external.pins import PINNED_CHAT_MODEL, PINNED_EMBED_MODEL
+from opmem_external.rates import (
+    RATE_CHAT_INPUT_PER_M,
+    RATE_CHAT_OUTPUT_PER_M,
+    RATE_EMBED_PER_M,
+    RATE_SOURCE_NOTE,
+    RATE_SOURCE_URL,
+)
+
+_ENC = tiktoken.get_encoding("cl100k_base")
+
+# Conservative max completion tokens per call type (upper bound, not typical).
+_MAX_OUTPUT_BY_KIND = {
+    "mem0_extract": 800,
+    "mem0_update": 400,
+    "letta_memory_step": 300,
+    "langmem_manage": 400,
+}
+
+_RETRY_MARGIN = 1.10  # settling / transient retries
+
+
+def _count(text: str) -> int:
+    return len(_ENC.encode(text or ""))
+
+
+def _count_json(obj: Any) -> int:
+    return _count(json.dumps(obj, ensure_ascii=False))
+
+
+@dataclass
+class BudgetLine:
+    lane: str
+    op: str
+    task: str
+    kind: str
+    input_tokens: int
+    output_tokens: int
+    embed_tokens: int
+
+    def chat_usd(self) -> float:
+        return (
+            self.input_tokens * RATE_CHAT_INPUT_PER_M / 1_000_000
+            + self.output_tokens * RATE_CHAT_OUTPUT_PER_M / 1_000_000
+        )
+
+    def embed_usd(self) -> float:
+        return self.embed_tokens * RATE_EMBED_PER_M / 1_000_000
+
+    def total_usd(self) -> float:
+        return self.chat_usd() + self.embed_usd()
+
+
+@dataclass
+class LaneBudget:
+    lane: str
+    lines: list[BudgetLine] = field(default_factory=list)
+    product_constraints: list[str] = field(default_factory=list)
+
+    def worst_case_usd(self) -> float:
+        base = sum(line.total_usd() for line in self.lines)
+        if self.lane == "zep":
+            return 0.0
+        return base * _RETRY_MARGIN
+
+    def token_totals(self) -> dict[str, int]:
+        return {
+            "chat_input_tokens": sum(l.input_tokens for l in self.lines),
+            "chat_output_tokens": sum(l.output_tokens for l in self.lines),
+            "embed_tokens": sum(l.embed_tokens for l in self.lines),
+        }
+
+    def to_dict(self) -> dict:
+        return {
+            "lane": self.lane,
+            "worst_case_usd": round(self.worst_case_usd(), 6),
+            "token_totals": self.token_totals(),
+            "line_count": len(self.lines),
+            "product_constraints": self.product_constraints,
+            "retry_margin": _RETRY_MARGIN,
+        }
+
+
+def _mem0_system_prompt() -> str:
+    from mem0.configs.prompts import ADDITIVE_EXTRACTION_PROMPT
+
+    return ADDITIVE_EXTRACTION_PROMPT
+
+
+def _mem0_extract_payload(content: str) -> list[dict]:
+    system = _mem0_system_prompt()
+    user = json.dumps(
+        {
+            "new_messages": [{"role": "user", "content": content}],
+            "summary": "",
+            "recently_extracted_memories": [],
+            "existing_memories": [],
+            "last_k_messages": [],
+            "observation_date": "2026-09-19",
+            "current_date": "2026-09-19",
+        },
+        ensure_ascii=False,
+    )
+    return [
+        {"role": "system", "content": system},
+        {"role": "user", "content": user},
+    ]
+
+
+def _letta_wrapped(content: str) -> dict:
+    return {
+        "model": PINNED_CHAT_MODEL,
+        "messages": [
+            {
+                "role": "system",
+                "content": "Extract durable user memory facts; memory-only lane (no answer generation).",
+            },
+            {"role": "user", "content": content},
+        ],
+        "max_tokens": _MAX_OUTPUT_BY_KIND["letta_memory_step"],
+    }
+
+
+def _langmem_manage_payload(content: str) -> dict:
+    return {
+        "model": PINNED_CHAT_MODEL,
+        "instructions": "Manage memory store entries for user facts (OpMem lane; no final answer).",
+        "input": content,
+    }
+
+
+def _zep_add_payload(content: str) -> dict:
+    return {
+        "messages": [
+            {
+                "role_type": "user",
+                "role": "opmem-user",
+                "content": content,
+            }
+        ],
+        "return_context": False,
+    }
+
+
+def build_lane_budgets(tasks: Iterable[dict]) -> dict[str, LaneBudget]:
+    budgets = {
+        "mem0-oss": LaneBudget("mem0-oss"),
+        "letta": LaneBudget("letta"),
+        "langmem": LaneBudget("langmem"),
+        "zep": LaneBudget("zep"),
+    }
+    budgets["zep"].product_constraints.append(
+        "Zep Cloud native extraction; OpenAI pins not applied; $0 paid ceiling."
+    )
+
+    for task in tasks:
+        name = task["name"]
+        for step in task["steps"]:
+            op = step["op"]
+            content = step.get("content", "")
+            query = step.get("query", "")
+
+            if op == "remember":
+                msgs = _mem0_extract_payload(content)
+                inp = sum(_count(m["content"]) for m in msgs)
+                budgets["mem0-oss"].lines.append(
+                    BudgetLine(
+                        "mem0-oss", op, name, "mem0_extract", inp,
+                        _MAX_OUTPUT_BY_KIND["mem0_extract"],
+                        _count(content),
+                    )
+                )
+                letta_p = _letta_wrapped(content)
+                budgets["letta"].lines.append(
+                    BudgetLine(
+                        "letta", op, name, "letta_memory", _count_json(letta_p),
+                        _MAX_OUTPUT_BY_KIND["letta_memory_step"],
+                        _count(content),
+                    )
+                )
+                lm = _langmem_manage_payload(content)
+                budgets["langmem"].lines.append(
+                    BudgetLine(
+                        "langmem", op, name, "langmem_manage", _count_json(lm),
+                        _MAX_OUTPUT_BY_KIND["langmem_manage"],
+                        _count(content),
+                    )
+                )
+                budgets["zep"].lines.append(
+                    BudgetLine("zep", op, name, "zep_add", _count_json(_zep_add_payload(content)), 0, 0)
+                )
+            elif op == "recall":
+                q = query or ""
+                for lane in ("mem0-oss", "letta", "langmem"):
+                    budgets[lane].lines.append(
+                        BudgetLine(lane, op, name, "embed_query", 0, 0, _count(q))
+                    )
+                # mem0 may re-embed on search path conservatively
+                budgets["mem0-oss"].lines.append(
+                    BudgetLine("mem0-oss", op, name, "embed_query_dup", 0, 0, _count(q))
+                )
+                budgets["zep"].lines.append(
+                    BudgetLine("zep", op, name, "zep_get", _count_json({"query": q}), 0, 0)
+                )
+            elif op == "revise":
+                msgs = _mem0_extract_payload(content)
+                inp = sum(_count(m["content"]) for m in msgs)
+                budgets["mem0-oss"].lines.append(
+                    BudgetLine(
+                        "mem0-oss", op, name, "mem0_update", inp,
+                        _MAX_OUTPUT_BY_KIND["mem0_update"],
+                        _count(content),
+                    )
+                )
+                budgets["letta"].lines.append(
+                    BudgetLine(
+                        "letta", op, name, "letta_update", _count_json(_letta_wrapped(content)),
+                        _MAX_OUTPUT_BY_KIND["letta_memory_step"],
+                        _count(content),
+                    )
+                )
+                budgets["langmem"].lines.append(
+                    BudgetLine(
+                        "langmem", op, name, "langmem_update", _count_json(_langmem_manage_payload(content)),
+                        _MAX_OUTPUT_BY_KIND["langmem_manage"],
+                        _count(content),
+                    )
+                )
+                budgets["zep"].lines.append(
+                    BudgetLine("zep", op, name, "zep_update", _count_json({"text": content}), 0, 0)
+                )
+            elif op == "forget":
+                for lane in ("mem0-oss", "letta", "langmem"):
+                    budgets[lane].lines.append(
+                        BudgetLine(lane, op, name, "delete_api", 50, 0, 0)
+                    )
+                budgets["zep"].lines.append(
+                    BudgetLine("zep", op, name, "zep_delete", 50, 0, 0)
+                )
+
+    return budgets
+
+
+def build_worksheet_from_budgets(
+    budgets: dict[str, LaneBudget],
+    *,
+    task_count: int,
+    step_count: int,
+) -> dict:
+    from opmem_external.pins import SPEND_CEILING_USD
+
+    lanes = []
+    openai_total = 0.0
+    for lane, budget in sorted(budgets.items()):
+        usd = budget.worst_case_usd()
+        openai_total += usd
+        lanes.append(
+            {
+                **budget.to_dict(),
+                "projected_worst_case_usd": round(usd, 6),
+                "ceiling_usd": SPEND_CEILING_USD.get(lane, 0.0),
+                "within_ceiling": usd <= SPEND_CEILING_USD.get(lane, 0.0),
+            }
+        )
+    lanes.append(
+        {
+            "lane": "_total_openai_worst_case",
+            "projected_worst_case_usd": round(openai_total, 6),
+            "ceiling_usd": SPEND_CEILING_USD["total"],
+            "within_ceiling": openai_total <= SPEND_CEILING_USD["total"],
+        }
+    )
+    return {
+        "mode": "payload_budget",
+        "task_count": task_count,
+        "step_count": step_count,
+        "pinned_chat_model": PINNED_CHAT_MODEL,
+        "pinned_embed_model": PINNED_EMBED_MODEL,
+        "rate_source_url": RATE_SOURCE_URL,
+        "rate_source_note": RATE_SOURCE_NOTE,
+        "ceilings_usd": SPEND_CEILING_USD,
+        "lanes": lanes,
+        "methodology": (
+            "tiktoken cl100k_base on serialized request bodies incl. mem0 ADDITIVE_EXTRACTION_PROMPT; "
+            "conservative max output caps; 1.10 retry margin; Zep excluded from OpenAI totals."
+        ),
+    }
+
+
+def load_tasks(fixture_dir: Path) -> list[dict]:
+    return [json.loads(p.read_text(encoding="utf-8")) for p in sorted(fixture_dir.glob("*.json"))]
diff --git a/evals/opmem_external/langmem_store.py b/evals/opmem_external/langmem_store.py
new file mode 100644
index 00000000..6773e0b8
--- /dev/null
+++ b/evals/opmem_external/langmem_store.py
@@ -0,0 +1,163 @@
+"""LangMem lane: local SDK + Postgres/pgvector (no InMemoryStore for measured runs)."""
+
+from __future__ import annotations
+
+import json
+import os
+import uuid
+from typing import Any
+
+from langchain_core.messages import HumanMessage
+
+from opmem_external.pins import PINNED_CHAT_MODEL, PINNED_EMBED_MODEL
+from opmem_external.protocol import OpMemLaneAdapter
+from opmem_external.telemetry import LaneTelemetry
+
+
+def _stringify_memory_value(val: Any) -> str:
+    if val is None:
+        return ""
+    if isinstance(val, str):
+        return val
+    if isinstance(val, dict):
+        for key in ("content", "text", "memory"):
+            inner = val.get(key)
+            if isinstance(inner, str):
+                return inner
+            if inner is not None:
+                return _stringify_memory_value(inner)
+        return json.dumps(val, ensure_ascii=False)
+    content = getattr(val, "content", None)
+    if isinstance(content, str):
+        return content
+    return str(val)
+
+
+class LangMemLaneAdapter(OpMemLaneAdapter):
+    name = "langmem"
+    write_settle_seconds = 0.5
+
+    def __init__(self, *, dry_run: bool = True, run_id: str | None = None) -> None:
+        super().__init__(dry_run=dry_run, run_id=run_id)
+        self._database_url = os.environ.get(
+            "DATABASE_URL",
+            "postgresql://opmem:opmem@127.0.0.1:5432/opmem_langmem",
+        )
+        if not dry_run and "InMemoryStore" in os.environ.get("LANGMEM_STORE", ""):
+            self._record_model_constraint("InMemoryStore is forbidden for measured OpMem runs")
+        self._store_ctx: Any = None
+        self._store: Any = None
+        self._manager: Any = None
+        self.telemetry = LaneTelemetry(self.name)
+        self._assert_pinned_models(chat_ok=True, embed_ok=True)
+
+    def lane_id(self) -> str:
+        return "langmem_postgres_pgvector"
+
+    def _user_id(self, actor: tuple[str, str]) -> str:
+        return self.namespace(actor)
+
+    def _ensure_store(self) -> Any:
+        if self._store is not None:
+            return self._store
+        from langgraph.store.postgres import PostgresStore
+
+        self._store_ctx = PostgresStore.from_conn_string(
+            self._database_url,
+            index={
+                "dims": 1536,
+                "embed": f"openai:{PINNED_EMBED_MODEL}",
+            },
+        )
+        self._store = self._store_ctx.__enter__()
+        self._store.setup()
+        return self._store
+
+    def _ensure_manager(self) -> Any:
+        if self._manager is not None:
+            return self._manager
+        from langmem import create_memory_store_manager
+
+        store = self._ensure_store()
+        self._manager = create_memory_store_manager(
+            f"openai:{PINNED_CHAT_MODEL}",
+            namespace=("opmem", "{langgraph_user_id}"),
+            store=store,
+            enable_deletes=True,
+            instructions=(
+                "Extract durable user memory facts only. Do not answer questions; "
+                "memory operations for OpMem benchmark lane."
+            ),
+        )
+        return self._manager
+
+    def _config(self, actor: tuple[str, str]) -> dict:
+        return {"configurable": {"langgraph_user_id": self._user_id(actor)}}
+
+    def _reset_remote(self) -> None:
+        self._manager = None
+        if self._store is not None:
+            user_prefix = self._user_id(("t1", "u1"))[:8]
+            try:
+                import psycopg
+
+                with psycopg.connect(self._database_url) as conn:
+                    with conn.cursor() as cur:
+                        cur.execute(
+                            "DELETE FROM store WHERE prefix LIKE %s",
+                            (f"opmem.%",),
+                        )
+                    conn.commit()
+            except Exception as exc:
+                self.telemetry.log_http("reset", 500, f"reset partial: {exc}")
+        self._memory_table.clear()
+
+    def _remember_live(self, actor: tuple[str, str], content: str, mem_id: str) -> list[str]:
+        manager = self._ensure_manager()
+        cfg = self._config(actor)
+        manager.invoke({"messages": [HumanMessage(content=content)]}, cfg)
+        self.telemetry.log_http("remember", 200, "langmem manager.invoke")
+        return [mem_id]
+
+    def _recall_live(self, actor: tuple[str, str], query: str) -> list[dict]:
+        store = self._ensure_store()
+        user = self._user_id(actor)
+        namespace = ("opmem", user)
+        hits = store.search(namespace, query=query, limit=10)
+        self.telemetry.log_http("recall", 200, str(hits)[:500])
+        out: list[dict] = []
+        for item in hits:
+            val = getattr(item, "value", item)
+            text = _stringify_memory_value(val)
+            key = getattr(item, "key", None) or str(uuid.uuid4())
+            out.append({"id": key, "content": text})
+        return out
+
+    def _forget_live(self, actor: tuple[str, str], memory_ids: list[str]) -> None:
+        store = self._ensure_store()
+        namespace = ("opmem", self._user_id(actor))
+        for mid in memory_ids:
+            try:
+                store.delete(namespace, mid)
+            except Exception:
+                pass
+        self.telemetry.log_http("forget", 200, f"deleted {memory_ids}")
+
+    def _revise_live(self, actor: tuple[str, str], memory_ids: list[str], content: str) -> None:
+        manager = self._ensure_manager()
+        cfg = self._config(actor)
+        manager.invoke(
+            {
+                "messages": [
+                    HumanMessage(
+                        content=f"Update memory to reflect: {content}"
+                    )
+                ]
+            },
+            cfg,
+        )
+        self.telemetry.log_http("revise", 200, "langmem revise invoke")
+        if self.write_settle_seconds:
+            import time
+
+            time.sleep(self.write_settle_seconds)
diff --git a/evals/opmem_external/letta_appserver.py b/evals/opmem_external/letta_appserver.py
new file mode 100644
index 00000000..e7d71def
--- /dev/null
+++ b/evals/opmem_external/letta_appserver.py
@@ -0,0 +1,137 @@
+"""Letta self-hosted App Server lane (pip server 0.11.x + archival block mapping)."""
+
+from __future__ import annotations
+
+import os
+import re
+from typing import Any
+
+from opmem_external.pins import PINNED_CHAT_MODEL, PINNED_EMBED_MODEL
+from opmem_external.protocol import OpMemLaneAdapter, _overlap
+from opmem_external.telemetry import LaneTelemetry
+
+_BLOCK_LABEL = "opmem_facts"
+_LINE_RE = re.compile(r"^\[(?P<id>[^\]]+)\]\s*(?P<body>.*)$")
+
+
+class LettaAppServerLaneAdapter(OpMemLaneAdapter):
+    name = "letta"
+    write_settle_seconds = 1.0
+
+    def __init__(self, *, dry_run: bool = True, run_id: str | None = None) -> None:
+        super().__init__(dry_run=dry_run, run_id=run_id)
+        self._base_url = os.environ.get("LETTA_BASE_URL", "http://127.0.0.1:8283").rstrip("/")
+        self._client: Any = None
+        self._agent_id: str | None = None
+        self.telemetry = LaneTelemetry(self.name)
+        self._assert_pinned_models(chat_ok=True, embed_ok=True)
+        self._record_model_constraint(
+            "Letta lane maps remember/revise/forget to agents.blocks.modify on "
+            f"'{_BLOCK_LABEL}' (line-per-memory). Recall uses lexical overlap, not agent chat."
+        )
+        if not dry_run and not os.environ.get("LETTA_APP_SERVER_TOKEN", "").strip():
+            # Local `letta server --no-secure` (pinned 0.11.7) does not require a token.
+            if not self._local_insecure_server():
+                self._missing_credentials.append("LETTA_APP_SERVER_TOKEN")
+
+    def lane_id(self) -> str:
+        return "letta_app_server_self_hosted"
+
+    def _local_insecure_server(self) -> bool:
+        host = self._base_url
+        return host.startswith("http://127.0.0.1") or host.startswith("http://localhost")
+
+    def _client_instance(self) -> Any:
+        if self._client is not None:
+            return self._client
+        from letta_client import Letta
+
+        token = os.environ.get("LETTA_APP_SERVER_TOKEN", "").strip()
+        kwargs: dict[str, Any] = {"base_url": self._base_url}
+        if token:
+            kwargs["api_key"] = token
+        self._client = Letta(**kwargs)
+        return self._client
+
+    def _model(self) -> str:
+        return f"openai/{PINNED_CHAT_MODEL}"
+
+    def _embedding(self) -> str:
+        return f"openai/{PINNED_EMBED_MODEL}"
+
+    def _reset_remote(self) -> None:
+        client = self._client_instance()
+        if self._agent_id:
+            try:
+                client.agents.delete(agent_id=self._agent_id)
+            except Exception:
+                pass
+        self._agent_id = None
+        ns = self.namespace(("t1", "u1"))
+        agent = client.agents.create(
+            name=f"opmem-{ns}",
+            memory_blocks=[
+                {"label": _BLOCK_LABEL, "value": ""},
+                {"label": "persona", "value": "OpMem external lane; memory block only (no chat answers)."},
+            ],
+            model=self._model(),
+            embedding=self._embedding(),
+        )
+        self._agent_id = agent.id
+        self.telemetry.log_http("reset", 200, f"agent {self._agent_id}")
+
+    def _read_lines(self) -> dict[str, str]:
+        client = self._client_instance()
+        if not self._agent_id:
+            self._reset_remote()
+        block = client.agents.blocks.retrieve(agent_id=self._agent_id, block_label=_BLOCK_LABEL)
+        text = block.value or ""
+        out: dict[str, str] = {}
+        for line in text.splitlines():
+            line = line.strip()
+            if not line:
+                continue
+            m = _LINE_RE.match(line)
+            if m:
+                out[m.group("id")] = m.group("body")
+        return out
+
+    def _write_lines(self, lines: dict[str, str]) -> None:
+        client = self._client_instance()
+        if not self._agent_id:
+            self._reset_remote()
+        body = "\n".join(f"[{mid}] {txt}" for mid, txt in sorted(lines.items()))
+        client.agents.blocks.modify(agent_id=self._agent_id, block_label=_BLOCK_LABEL, value=body)
+        self.telemetry.log_http("blocks.modify", 200, f"{len(lines)} facts")
+
+    def _remember_live(self, actor: tuple[str, str], content: str, mem_id: str) -> list[str]:
+        lines = self._read_lines()
+        lines[mem_id] = content
+        self._write_lines(lines)
+        return [mem_id]
+
+    def _recall_live(self, actor: tuple[str, str], query: str) -> list[dict]:
+        lines = self._read_lines()
+        scored = sorted(lines.items(), key=lambda item: -_overlap(query, item[1]))
+        out = [{"id": mid, "content": txt} for mid, txt in scored if _overlap(query, txt) > 0]
+        self.telemetry.log_http("recall", 200, str(out)[:500])
+        return out
+
+    def _forget_live(self, actor: tuple[str, str], memory_ids: list[str]) -> None:
+        lines = self._read_lines()
+        for mid in memory_ids:
+            lines.pop(mid, None)
+        self._write_lines(lines)
+        self.telemetry.log_http("forget", 200, f"removed {memory_ids}")
+
+    def _revise_live(self, actor: tuple[str, str], memory_ids: list[str], content: str) -> None:
+        if not memory_ids:
+            return
+        lines = self._read_lines()
+        lines[memory_ids[0]] = content
+        self._write_lines(lines)
+        self.telemetry.log_http("revise", 200, f"updated {memory_ids[0]}")
+        if self.write_settle_seconds:
+            import time
+
+            time.sleep(self.write_settle_seconds)
diff --git a/evals/opmem_external/mem0_oss.py b/evals/opmem_external/mem0_oss.py
new file mode 100644
index 00000000..75ed4b52
--- /dev/null
+++ b/evals/opmem_external/mem0_oss.py
@@ -0,0 +1,126 @@
+"""Mem0 OSS lane: Python package + local Qdrant (no Platform API key)."""
+
+from __future__ import annotations
+
+import os
+import shutil
+from pathlib import Path
+from typing import Any
+
+from opmem_external.pins import PINNED_CHAT_MODEL, PINNED_EMBED_MODEL
+from opmem_external.protocol import OpMemLaneAdapter
+from opmem_external.telemetry import LaneTelemetry
+
+
+class Mem0OSSLaneAdapter(OpMemLaneAdapter):
+    name = "mem0-oss"
+    write_settle_seconds = 2.0
+
+    def __init__(self, *, dry_run: bool = True, run_id: str | None = None) -> None:
+        super().__init__(dry_run=dry_run, run_id=run_id)
+        self._qdrant_root = Path(os.environ.get("OPMEM_QDRANT_ROOT", "/tmp/opmem-qdrant"))
+        self._memory: Any = None
+        self._task_store_path: Path | None = None
+        self.telemetry = LaneTelemetry(self.name)
+        self._assert_pinned_models(chat_ok=True, embed_ok=True)
+
+    def lane_id(self) -> str:
+        return "mem0_oss_local_qdrant"
+
+    def _config(self, store_path: Path, collection: str) -> dict:
+        return {
+            "vector_store": {
+                "provider": "qdrant",
+                "config": {
+                    "path": str(store_path),
+                    "collection_name": collection,
+                    "embedding_model_dims": 1536,
+                    "on_disk": True,
+                },
+            },
+            "llm": {
+                "provider": "openai",
+                "config": {"model": PINNED_CHAT_MODEL, "temperature": 0},
+            },
+            "embedder": {
+                "provider": "openai",
+                "config": {"model": PINNED_EMBED_MODEL},
+            },
+        }
+
+    def _load_memory(self) -> Any:
+        if self._memory is not None:
+            return self._memory
+        from mem0 import Memory
+
+        store_path = self._task_store_path or (self._qdrant_root / "default")
+        store_path.mkdir(parents=True, exist_ok=True)
+        collection = f"opmem_{self._task_name or 'task'}"
+        self._memory = Memory.from_config(self._config(store_path, collection))
+        return self._memory
+
+    def _reset_remote(self) -> None:
+        self._memory = None
+        path = self._qdrant_root / self._run_id / self._task_name
+        if path.exists():
+            shutil.rmtree(path, ignore_errors=True)
+        self._task_store_path = path
+        self._task_store_path.mkdir(parents=True, exist_ok=True)
+
+    def _user_id(self, actor: tuple[str, str]) -> str:
+        return self.namespace(actor)
+
+    def _remember_live(self, actor: tuple[str, str], content: str, mem_id: str) -> list[str]:
+        mem = self._load_memory()
+        user_id = self._user_id(actor)
+        result = mem.add(
+            [{"role": "user", "content": content}],
+            user_id=user_id,
+            infer=True,
+        )
+        self.telemetry.log_http("remember", 200, str(result)[:500])
+        ids = []
+        for item in (result or {}).get("results", []) if isinstance(result, dict) else []:
+            if isinstance(item, dict) and item.get("id"):
+                ids.append(item["id"])
+        if not ids and isinstance(result, dict):
+            for item in result.get("results", []):
+                mid = item.get("memory_id") or item.get("id")
+                if mid:
+                    ids.append(mid)
+        return ids[-1:] if ids else [mem_id]
+
+    def _recall_live(self, actor: tuple[str, str], query: str) -> list[dict]:
+        mem = self._load_memory()
+        user_id = self._user_id(actor)
+        raw = mem.search(query, filters={"user_id": user_id}, top_k=10)
+        self.telemetry.log_http("recall", 200, str(raw)[:500])
+        out = []
+        for item in raw.get("results", []) if isinstance(raw, dict) else raw or []:
+            if not isinstance(item, dict):
+                continue
+            raw = item.get("memory") or item.get("text") or item.get("content") or ""
+            if isinstance(raw, dict):
+                text = raw.get("data") or raw.get("text") or str(raw)
+            else:
+                text = str(raw)
+            mid = item.get("id") or item.get("memory_id") or ""
+            out.append({"id": mid, "content": text})
+        return out
+
+    def _forget_live(self, actor: tuple[str, str], memory_ids: list[str]) -> None:
+        mem = self._load_memory()
+        for mid in memory_ids:
+            mem.delete(mid)
+            self.telemetry.log_http("forget", 200, f"deleted {mid}")
+
+    def _revise_live(self, actor: tuple[str, str], memory_ids: list[str], content: str) -> None:
+        mem = self._load_memory()
+        if not memory_ids:
+            return
+        mem.update(memory_ids[0], data=content)
+        self.telemetry.log_http("revise", 200, f"updated {memory_ids[0]}")
+        if self.write_settle_seconds:
+            import time
+
+            time.sleep(self.write_settle_seconds)
diff --git a/evals/opmem_external/pins.py b/evals/opmem_external/pins.py
new file mode 100644
index 00000000..d715ff31
--- /dev/null
+++ b/evals/opmem_external/pins.py
@@ -0,0 +1,28 @@
+"""Pinned models and spend ceilings for OpMem external lane comparisons."""
+
+from __future__ import annotations
+
+PINNED_CHAT_MODEL = "gpt-4.1-mini-2025-04-14"
+PINNED_EMBED_MODEL = "text-embedding-3-small"
+
+# Official list prices (USD per 1M tokens) for worksheet generation — 2025-04 OpenAI public rates.
+RATE_CHAT_INPUT_PER_M = 0.40
+RATE_CHAT_OUTPUT_PER_M = 1.60
+RATE_EMBED_PER_M = 0.02
+
+# Owner-approved spend authority (USD); live runs must stay at or below projected totals.
+SPEND_CEILING_USD: dict[str, float] = {
+    "mem0-oss": 1.0,
+    "letta": 1.0,
+    "langmem": 1.0,
+    "zep": 0.0,
+    "total": 3.0,
+}
+
+LANE_ENV_REQUIRED: dict[str, tuple[str, ...]] = {
+    "mem0-oss": ("OPENAI_API_KEY",),  # local Qdrant; no MEM0_API_KEY
+    "zep": ("ZEP_API_KEY",),
+    # LETTA_APP_SERVER_TOKEN optional for local `letta server --no-secure` (see external-lanes-setup.md).
+    "letta": ("OPENAI_API_KEY",),
+    "langmem": ("OPENAI_API_KEY", "DATABASE_URL"),
+}
diff --git a/evals/opmem_external/protocol.py b/evals/opmem_external/protocol.py
new file mode 100644
index 00000000..20de49ca
--- /dev/null
+++ b/evals/opmem_external/protocol.py
@@ -0,0 +1,220 @@
+"""Shared OpMem external-lane adapter protocol."""
+
+from __future__ import annotations
+
+import hashlib
+import json
+import os
+import time
+from abc import ABC, abstractmethod
+from dataclasses import dataclass, field
+from pathlib import Path
+from typing import Any
+
+from opmem_external.pins import LANE_ENV_REQUIRED, PINNED_CHAT_MODEL, PINNED_EMBED_MODEL
+from opmem_external.token_estimate import LaneDryRunStats, OperationCounts, TokenLedger
+
+
+@dataclass
+class AdapterResponseLog:
+    lane: str
+    task: str
+    op: str
+    status: int
+    body_preview: str
+    redacted: bool = True
+
+
+class OpMemLaneAdapter(ABC):
+    """Maps neutral remember/recall/revise/forget onto one external memory lane."""
+
+    name: str
+    supports_pinned_models: bool = True
+    write_settle_seconds: float = 0.0
+
+    def __init__(self, *, dry_run: bool = True, run_id: str | None = None) -> None:
+        self.dry_run = dry_run
+        self._run_id = run_id or os.environ.get("OPMEM_RUN_ID", "opmem-pin-0001")
+        self._task_name = ""
+        self._stats = LaneDryRunStats(lane=self.name)
+        self._memory_table: dict[str, dict[str, str]] = {}
+        self._response_logs: list[AdapterResponseLog] = []
+        self._missing_credentials = self._check_credentials()
+        self._stats.missing_credentials = list(self._missing_credentials)
+        self._stats.credentials_present = not self._missing_credentials
+
+    @abstractmethod
+    def lane_id(self) -> str:
+        ...
+
+    def _check_credentials(self) -> list[str]:
+        missing = []
+        for key in LANE_ENV_REQUIRED.get(self.name, ()):
+            if not os.environ.get(key, "").strip():
+                missing.append(key)
+        return missing
+
+    def available(self) -> bool:
+        if self.dry_run:
+            return True
+        return not self._missing_credentials
+
+    def begin_task(self, task_name: str) -> None:
+        self._task_name = task_name
+        self.reset_namespace()
+
+    def namespace(self, actor: tuple[str, str]) -> str:
+        tenant, subject = actor
+        raw = f"{self._run_id}|{self.name}|{self._task_name}|{tenant}|{subject}"
+        digest = hashlib.sha256(raw.encode("utf-8")).hexdigest()[:16]
+        return f"opmem-{digest}"
+
+    def reset_namespace(self) -> None:
+        self._stats.ops.reset += 1
+        keys = [k for k in self._memory_table if k.startswith(f"{self._task_name}:")]
+        for k in keys:
+            del self._memory_table[k]
+        if not self.dry_run:
+            self._reset_remote()
+
+    @abstractmethod
+    def _reset_remote(self) -> None:
+        ...
+
+    def remember(self, actor: tuple[str, str], content: str) -> list[str]:
+        self._stats.ops.remember += 1
+        ns = self._table_key(actor)
+        mem_id = self._deterministic_id("remember", content)
+        if self.dry_run:
+            self._simulate_remember(content)
+            self._memory_table.setdefault(ns, {})[mem_id] = content
+            self._log_response("remember", 200, {"memory_id": mem_id, "dry_run": True})
+            return [mem_id]
+        return self._remember_live(actor, content, mem_id)
+
+    def recall(self, actor: tuple[str, str], query: str) -> list[dict]:
+        self._stats.ops.recall += 1
+        if self.dry_run:
+            self._simulate_recall(query)
+            ns = self._table_key(actor)
+            rows = self._memory_table.get(ns, {})
+            scored = sorted(rows.items(), key=lambda item: -_overlap(query, item[1]))
+            out = [{"id": mid, "content": txt} for mid, txt in scored if _overlap(query, txt) > 0]
+            self._log_response("recall", 200, {"results": out, "dry_run": True})
+            return out
+        return self._recall_live(actor, query)
+
+    def forget(self, actor: tuple[str, str], memory_ids: list[str]) -> None:
+        self._stats.ops.forget += 1
+        if self.dry_run:
+            ns = self._table_key(actor)
+            bucket = self._memory_table.get(ns, {})
+            for mid in memory_ids:
+                bucket.pop(mid, None)
+            self._simulate_forget()
+            self._log_response("forget", 200, {"deleted": memory_ids, "dry_run": True})
+            return
+        self._forget_live(actor, memory_ids)
+
+    def revise(self, actor: tuple[str, str], memory_ids: list[str], content: str) -> None:
+        self._stats.ops.revise += 1
+        if self.dry_run:
+            ns = self._table_key(actor)
+            bucket = self._memory_table.get(ns, {})
+            if memory_ids:
+                bucket[memory_ids[0]] = content
+            self._simulate_revise(content)
+            self._log_response("revise", 200, {"memory_id": memory_ids[:1], "dry_run": True})
+            if self.write_settle_seconds:
+                time.sleep(self.write_settle_seconds)
+            return
+        self._revise_live(actor, memory_ids, content)
+
+    def _table_key(self, actor: tuple[str, str]) -> str:
+        return f"{self._task_name}:{self.namespace(actor)}"
+
+    def _deterministic_id(self, op: str, content: str) -> str:
+        seed = f"{self._run_id}:{self._task_name}:{op}:{content}"
+        return "m-" + hashlib.sha256(seed.encode("utf-8")).hexdigest()[:12]
+
+    def _log_response(self, op: str, status: int, body: Any) -> None:
+        preview = json.dumps(body, ensure_ascii=False)[:500]
+        self._response_logs.append(
+            AdapterResponseLog(
+                lane=self.name,
+                task=self._task_name,
+                op=op,
+                status=status,
+                body_preview=preview,
+            )
+        )
+        self._stats.ops.http_calls += 1
+
+    def _record_model_constraint(self, reason: str) -> None:
+        if reason not in self._stats.product_constraints:
+            self._stats.product_constraints.append(reason)
+
+    def _assert_pinned_models(self, chat_ok: bool, embed_ok: bool) -> None:
+        if not chat_ok:
+            self._record_model_constraint(
+                f"lane {self.name} cannot pin chat model {PINNED_CHAT_MODEL}"
+            )
+        if not embed_ok:
+            self._record_model_constraint(
+                f"lane {self.name} cannot pin embed model {PINNED_EMBED_MODEL}"
+            )
+
+    def _simulate_remember(self, content: str) -> None:
+        self._stats.tokens.add_chat(
+            f"[extract] model={PINNED_CHAT_MODEL}\n{content}",
+            '{"facts":[]}',
+        )
+        self._stats.tokens.add_embed(content)
+
+    def _simulate_recall(self, query: str) -> None:
+        self._stats.tokens.add_embed(query)
+
+    def _simulate_revise(self, content: str) -> None:
+        self._stats.tokens.add_chat(f"[update] model={PINNED_CHAT_MODEL}\n{content}", "{}")
+
+    def _simulate_forget(self) -> None:
+        pass
+
+    @abstractmethod
+    def _remember_live(self, actor: tuple[str, str], content: str, mem_id: str) -> list[str]:
+        ...
+
+    @abstractmethod
+    def _recall_live(self, actor: tuple[str, str], query: str) -> list[dict]:
+        ...
+
+    @abstractmethod
+    def _forget_live(self, actor: tuple[str, str], memory_ids: list[str]) -> None:
+        ...
+
+    @abstractmethod
+    def _revise_live(self, actor: tuple[str, str], memory_ids: list[str], content: str) -> None:
+        ...
+
+    def dry_run_stats(self) -> LaneDryRunStats:
+        return self._stats
+
+    def write_raw_log(self, path: Path) -> None:
+        path.parent.mkdir(parents=True, exist_ok=True)
+        rows = [
+            {
+                "lane": r.lane,
+                "task": r.task,
+                "op": r.op,
+                "status": r.status,
+                "body_preview": r.body_preview,
+            }
+            for r in self._response_logs
+        ]
+        path.write_text(json.dumps(rows, indent=2) + "\n", encoding="utf-8")
+
+
+def _overlap(query: str, content: str) -> int:
+    q = {t for t in query.lower().split() if t}
+    c = {t for t in content.lower().split() if t}
+    return len(q & c)
diff --git a/evals/opmem_external/rates.py b/evals/opmem_external/rates.py
new file mode 100644
index 00000000..66043bb7
--- /dev/null
+++ b/evals/opmem_external/rates.py
@@ -0,0 +1,28 @@
+"""OpenAI list prices used for OpMem external lane worksheets (verify before live runs)."""
+
+from __future__ import annotations
+
+# Source: https://openai.com/api/pricing/ (GPT-4.1 mini + text-embedding-3-small), captured 2026-09-19.
+RATE_SOURCE_URL = "https://openai.com/api/pricing/"
+RATE_SOURCE_NOTE = (
+    "gpt-4.1-mini-2025-04-14: $0.40/1M input, $1.60/1M output; "
+    "text-embedding-3-small: $0.02/1M tokens"
+)
+
+from opmem_external.pins import (  # noqa: E402
+    PINNED_CHAT_MODEL,
+    PINNED_EMBED_MODEL,
+    RATE_CHAT_INPUT_PER_M,
+    RATE_CHAT_OUTPUT_PER_M,
+    RATE_EMBED_PER_M,
+)
+
+__all__ = [
+    "RATE_SOURCE_URL",
+    "RATE_SOURCE_NOTE",
+    "PINNED_CHAT_MODEL",
+    "PINNED_EMBED_MODEL",
+    "RATE_CHAT_INPUT_PER_M",
+    "RATE_CHAT_OUTPUT_PER_M",
+    "RATE_EMBED_PER_M",
+]
diff --git a/evals/opmem_external/requirements-external-lanes.txt b/evals/opmem_external/requirements-external-lanes.txt
new file mode 100644
index 00000000..f0eaafe8
--- /dev/null
+++ b/evals/opmem_external/requirements-external-lanes.txt
@@ -0,0 +1,10 @@
+# Pinned deps for OpMem external lane reproduction (install in eval venv).
+mem0ai==2.1.0
+qdrant-client==1.19.1
+tiktoken==0.14.0
+langmem==0.0.30
+langgraph-checkpoint-postgres==3.1.2
+zep-cloud==3.28.0
+letta==0.11.7
+letta-client==0.1.307
+sqlite-vec
diff --git a/evals/opmem_external/schemas/lane-manifest.schema.json b/evals/opmem_external/schemas/lane-manifest.schema.json
new file mode 100644
index 00000000..a92db3a1
--- /dev/null
+++ b/evals/opmem_external/schemas/lane-manifest.schema.json
@@ -0,0 +1,15 @@
+{
+  "$schema": "https://json-schema.org/draft/2020-12/schema",
+  "title": "OpMemExternalLaneManifest",
+  "type": "object",
+  "required": ["lane", "pinned_chat_model", "pinned_embed_model", "namespace_scheme"],
+  "properties": {
+    "lane": { "type": "string", "enum": ["mem0-oss", "zep", "letta", "langmem"] },
+    "pinned_chat_model": { "const": "gpt-4.1-mini-2025-04-14" },
+    "pinned_embed_model": { "const": "text-embedding-3-small" },
+    "namespace_scheme": { "const": "opmem-sha256(run_id|lane|task|tenant|subject)[:16]" },
+    "reset_per_task": { "const": true },
+    "dry_run_default": { "const": true },
+    "spend_ceiling_usd": { "type": "number" }
+  }
+}
diff --git a/evals/opmem_external/telemetry.py b/evals/opmem_external/telemetry.py
new file mode 100644
index 00000000..ba3b9f25
--- /dev/null
+++ b/evals/opmem_external/telemetry.py
@@ -0,0 +1,53 @@
+"""Aggregate OpenAI usage for external lane runs (secrets redacted in logs)."""
+
+from __future__ import annotations
+
+import json
+import re
+from dataclasses import dataclass, field
+from pathlib import Path
+
+_SECRET_RE = re.compile(r"(sk-[A-Za-z0-9_-]{10,}|api[_-]?key[\"']?\s*[:=]\s*[\"'][^\"']+)", re.I)
+
+
+@dataclass
+class UsageTotals:
+    chat_input_tokens: int = 0
+    chat_output_tokens: int = 0
+    embed_tokens: int = 0
+    estimated_usd: float = 0.0
+    http_requests: int = 0
+
+    def to_dict(self) -> dict:
+        return {
+            "chat_input_tokens": self.chat_input_tokens,
+            "chat_output_tokens": self.chat_output_tokens,
+            "embed_tokens": self.embed_tokens,
+            "estimated_usd": round(self.estimated_usd, 6),
+            "http_requests": self.http_requests,
+        }
+
+
+@dataclass
+class LaneTelemetry:
+    lane: str
+    usage: UsageTotals = field(default_factory=UsageTotals)
+    events: list[dict] = field(default_factory=list)
+
+    def add_openai_usage(self, *, input_tokens: int = 0, output_tokens: int = 0, embed_tokens: int = 0, usd: float = 0.0) -> None:
+        self.usage.chat_input_tokens += input_tokens
+        self.usage.chat_output_tokens += output_tokens
+        self.usage.embed_tokens += embed_tokens
+        self.usage.estimated_usd += usd
+
+    def log_http(self, op: str, status: int, body_preview: str) -> None:
+        self.usage.http_requests += 1
+        redacted = _SECRET_RE.sub("***REDACTED***", body_preview)[:500]
+        self.events.append({"op": op, "status": status, "body_preview": redacted})
+
+    def write(self, path: Path) -> None:
+        path.parent.mkdir(parents=True, exist_ok=True)
+        path.write_text(
+            json.dumps({"lane": self.lane, "usage": self.usage.to_dict(), "events": self.events}, indent=2) + "\n",
+            encoding="utf-8",
+        )
diff --git a/evals/opmem_external/test_lane_contract.py b/evals/opmem_external/test_lane_contract.py
new file mode 100644
index 00000000..9d387696
--- /dev/null
+++ b/evals/opmem_external/test_lane_contract.py
@@ -0,0 +1,98 @@
+#!/usr/bin/env python3
+"""Contract tests for OpMem external lanes (no live vendor APIs)."""
+from __future__ import annotations
+
+import json
+import os
+import unittest
+from pathlib import Path
+from unittest import mock
+
+# Ensure evals/ is on path when run as module or script
+import sys
+
+ROOT = Path(__file__).resolve().parents[1]
+sys.path.insert(0, str(ROOT))
+
+from opmem_external import LANE_ADAPTERS  # noqa: E402
+from opmem_external.pins import SPEND_CEILING_USD  # noqa: E402
+from opmem_external.worksheet import build_worksheet  # noqa: E402
+from run_opmem import run_task  # noqa: E402
+
+
+class LaneContractTests(unittest.TestCase):
+    def setUp(self) -> None:
+        self._env = mock.patch.dict(os.environ, {}, clear=False)
+        self._env.start()
+        os.environ.pop("OPENAI_API_KEY", None)
+        os.environ.pop("ZEP_API_KEY", None)
+        os.environ.pop("LETTA_APP_SERVER_TOKEN", None)
+        os.environ.pop("DATABASE_URL", None)
+
+    def tearDown(self) -> None:
+        self._env.stop()
+
+    def test_namespace_is_deterministic(self) -> None:
+        adapter = LANE_ADAPTERS["mem0-oss"](dry_run=True, run_id="fixed-run")
+        adapter.begin_task("cor01_basic_revision")
+        ns1 = adapter.namespace(("t1", "u1"))
+        adapter2 = LANE_ADAPTERS["mem0-oss"](dry_run=True, run_id="fixed-run")
+        adapter2.begin_task("cor01_basic_revision")
+        ns2 = adapter2.namespace(("t1", "u1"))
+        self.assertEqual(ns1, ns2)
+
+    def test_dry_run_remember_recall_no_credentials(self) -> None:
+        adapter = LANE_ADAPTERS["langmem"](dry_run=True, run_id="r1")
+        self.assertTrue(adapter.available())
+        adapter.begin_task("sup01_basic_forget")
+        ids = adapter.remember(("t1", "u1"), "door code is 4455")
+        self.assertEqual(len(ids), 1)
+        hits = adapter.recall(("t1", "u1"), "door code")
+        self.assertGreaterEqual(len(hits), 1)
+
+    def test_zep_records_model_constraint(self) -> None:
+        adapter = LANE_ADAPTERS["zep"](dry_run=True)
+        stats = adapter.dry_run_stats()
+        self.assertTrue(stats.product_constraints)
+
+    def test_worksheet_within_ceilings_dry_run(self) -> None:
+        fixture_dir = ROOT.parent / "fixtures" / "opmem"
+        tasks = [json.loads(p.read_text()) for p in sorted(fixture_dir.glob("*.json"))]
+        stats = []
+        for name, factory in LANE_ADAPTERS.items():
+            adapter = factory(dry_run=True, run_id="worksheet-test")
+            for task in tasks:
+                run_task(adapter, task)
+            stats.append(adapter.dry_run_stats())
+        ws = build_worksheet(stats, task_count=len(tasks), step_count=47)
+        total = next(row for row in ws.lanes if row["lane"] == "_total_openai_estimate")
+        self.assertLessEqual(total["projected_usd"], SPEND_CEILING_USD["total"])
+        for row in ws.lanes:
+            if row["lane"] in SPEND_CEILING_USD and not row["lane"].startswith("_"):
+                self.assertLessEqual(row["projected_usd"], SPEND_CEILING_USD[row["lane"]])
+
+    def test_raw_log_does_not_contain_env_secrets(self) -> None:
+        with mock.patch.dict(
+            os.environ,
+            {"OPENAI_API_KEY": "sk-test-secret-value-12345", "ZEP_API_KEY": "zep-secret"},
+            clear=False,
+        ):
+            adapter = LANE_ADAPTERS["mem0-oss"](dry_run=True, run_id="log-test")
+            adapter.begin_task("dup01_idempotent_remember")
+            adapter.remember(("t1", "u1"), "hello")
+            log_path = Path(self._testMethodName + ".json")
+            try:
+                adapter.write_raw_log(log_path)
+                text = log_path.read_text()
+                self.assertNotIn("sk-test-secret", text)
+                self.assertNotIn("zep-secret", text)
+            finally:
+                log_path.unlink(missing_ok=True)
+
+    def test_live_mode_requires_credentials(self) -> None:
+        adapter = LANE_ADAPTERS["letta"](dry_run=False)
+        self.assertFalse(adapter.available())
+
+
+if __name__ == "__main__":
+    unittest.main()
diff --git a/evals/opmem_external/token_estimate.py b/evals/opmem_external/token_estimate.py
new file mode 100644
index 00000000..441dc3f5
--- /dev/null
+++ b/evals/opmem_external/token_estimate.py
@@ -0,0 +1,94 @@
+"""Dry-run token estimation (no vendor calls)."""
+
+from __future__ import annotations
+
+from dataclasses import dataclass, field
+
+from opmem_external.pins import (
+    PINNED_CHAT_MODEL,
+    PINNED_EMBED_MODEL,
+    RATE_CHAT_INPUT_PER_M,
+    RATE_CHAT_OUTPUT_PER_M,
+    RATE_EMBED_PER_M,
+)
+
+
+def estimate_tokens(text: str) -> int:
+    """Conservative heuristic when tiktoken is not a repo dependency."""
+    stripped = (text or "").strip()
+    if not stripped:
+        return 0
+    return max(1, len(stripped) // 4)
+
+
+@dataclass
+class TokenLedger:
+    chat_input_tokens: int = 0
+    chat_output_tokens: int = 0
+    embed_tokens: int = 0
+    chat_model: str = PINNED_CHAT_MODEL
+    embed_model: str = PINNED_EMBED_MODEL
+
+    def add_chat(self, prompt: str, completion: str = "") -> None:
+        self.chat_input_tokens += estimate_tokens(prompt)
+        self.chat_output_tokens += estimate_tokens(completion)
+
+    def add_embed(self, text: str) -> None:
+        self.embed_tokens += estimate_tokens(text)
+
+    def estimated_usd(self) -> float:
+        return (
+            self.chat_input_tokens * RATE_CHAT_INPUT_PER_M / 1_000_000
+            + self.chat_output_tokens * RATE_CHAT_OUTPUT_PER_M / 1_000_000
+            + self.embed_tokens * RATE_EMBED_PER_M / 1_000_000
+        )
+
+    def to_dict(self) -> dict:
+        return {
+            "chat_model": self.chat_model,
+            "embed_model": self.embed_model,
+            "chat_input_tokens": self.chat_input_tokens,
+            "chat_output_tokens": self.chat_output_tokens,
+            "embed_tokens": self.embed_tokens,
+            "estimated_usd": round(self.estimated_usd(), 6),
+        }
+
+
+@dataclass
+class OperationCounts:
+    remember: int = 0
+    recall: int = 0
+    revise: int = 0
+    forget: int = 0
+    reset: int = 0
+    http_calls: int = 0
+
+    def to_dict(self) -> dict:
+        return {
+            "remember": self.remember,
+            "recall": self.recall,
+            "revise": self.revise,
+            "forget": self.forget,
+            "reset": self.reset,
+            "http_calls": self.http_calls,
+        }
+
+
+@dataclass
+class LaneDryRunStats:
+    lane: str
+    ops: OperationCounts = field(default_factory=OperationCounts)
+    tokens: TokenLedger = field(default_factory=TokenLedger)
+    product_constraints: list[str] = field(default_factory=list)
+    credentials_present: bool = False
+    missing_credentials: list[str] = field(default_factory=list)
+
+    def to_dict(self) -> dict:
+        return {
+            "lane": self.lane,
+            "operations": self.ops.to_dict(),
+            "tokens": self.tokens.to_dict(),
+            "product_constraints": self.product_constraints,
+            "credentials_present": self.credentials_present,
+            "missing_credentials": self.missing_credentials,
+        }
diff --git a/evals/opmem_external/worksheet.py b/evals/opmem_external/worksheet.py
new file mode 100644
index 00000000..54f8131a
--- /dev/null
+++ b/evals/opmem_external/worksheet.py
@@ -0,0 +1,97 @@
+"""Per-lane cost worksheet from dry-run instrumentation."""
+
+from __future__ import annotations
+
+import json
+from dataclasses import dataclass
+from datetime import datetime, timezone
+from pathlib import Path
+
+from opmem_external.pins import (
+    PINNED_CHAT_MODEL,
+    PINNED_EMBED_MODEL,
+    RATE_CHAT_INPUT_PER_M,
+    RATE_CHAT_OUTPUT_PER_M,
+    RATE_EMBED_PER_M,
+    SPEND_CEILING_USD,
+)
+from opmem_external.token_estimate import LaneDryRunStats
+
+
+@dataclass
+class Worksheet:
+    generated_utc: str
+    task_count: int
+    step_count: int
+    rates: dict
+    pinned_models: dict
+    ceilings_usd: dict
+    lanes: list[dict]
+
+    def to_dict(self) -> dict:
+        return {
+            "generated_utc": self.generated_utc,
+            "task_count": self.task_count,
+            "step_count": self.step_count,
+            "rates": self.rates,
+            "pinned_models": self.pinned_models,
+            "ceilings_usd": self.ceilings_usd,
+            "lanes": self.lanes,
+        }
+
+
+def build_worksheet(
+    lane_stats: list[LaneDryRunStats],
+    *,
+    task_count: int,
+    step_count: int,
+) -> Worksheet:
+    lanes_out = []
+    total_usd = 0.0
+    for stats in lane_stats:
+        usd = stats.tokens.estimated_usd()
+        if stats.lane == "zep":
+            usd = 0.0
+        total_usd += usd
+        ceiling = SPEND_CEILING_USD.get(stats.lane, 0.0)
+        lanes_out.append(
+            {
+                **stats.to_dict(),
+                "projected_usd": round(usd, 6),
+                "ceiling_usd": ceiling,
+                "within_ceiling": usd <= ceiling,
+                "live_run_allowed": False,
+                "live_run_note": (
+                    "Phase 1 dry-run only; enable after credentials present and "
+                    "measured estimate confirmed under ceiling."
+                ),
+            }
+        )
+    lanes_out.append(
+        {
+            "lane": "_total_openai_estimate",
+            "projected_usd": round(total_usd, 6),
+            "ceiling_usd": SPEND_CEILING_USD["total"],
+            "within_ceiling": total_usd <= SPEND_CEILING_USD["total"],
+        }
+    )
+    return Worksheet(
+        generated_utc=datetime.now(timezone.utc).strftime("%Y-%m-%dT%H:%M:%SZ"),
+        task_count=task_count,
+        step_count=step_count,
+        rates={
+            PINNED_CHAT_MODEL: {
+                "input_per_m": RATE_CHAT_INPUT_PER_M,
+                "output_per_m": RATE_CHAT_OUTPUT_PER_M,
+            },
+            PINNED_EMBED_MODEL: {"per_m": RATE_EMBED_PER_M},
+        },
+        pinned_models={"chat": PINNED_CHAT_MODEL, "embed": PINNED_EMBED_MODEL},
+        ceilings_usd=SPEND_CEILING_USD,
+        lanes=lanes_out,
+    )
+
+
+def write_worksheet(path: Path, worksheet: Worksheet) -> None:
+    path.parent.mkdir(parents=True, exist_ok=True)
+    path.write_text(json.dumps(worksheet.to_dict(), indent=2) + "\n", encoding="utf-8")
diff --git a/evals/opmem_external/zep_cloud.py b/evals/opmem_external/zep_cloud.py
new file mode 100644
index 00000000..03df6abe
--- /dev/null
+++ b/evals/opmem_external/zep_cloud.py
@@ -0,0 +1,162 @@
+"""Zep Cloud lane (free tier only; $0 paid ceiling). SDK 3.x thread + graph API."""
+
+from __future__ import annotations
+
+import os
+import re
+
+from opmem_external.protocol import OpMemLaneAdapter, _overlap
+from opmem_external.telemetry import LaneTelemetry
+
+_BILLING_HINT = re.compile(r"(upgrade|payment|billing|subscribe|credit|quota|limit exceeded)", re.I)
+
+
+class ZepCloudLaneAdapter(OpMemLaneAdapter):
+    name = "zep"
+    write_settle_seconds = 1.5
+
+    def __init__(self, *, dry_run: bool = True, run_id: str | None = None) -> None:
+        super().__init__(dry_run=dry_run, run_id=run_id)
+        self._api_key = os.environ.get("ZEP_API_KEY", "")
+        self.telemetry = LaneTelemetry(self.name)
+        self._thread_id: str | None = None
+        self._user_id: str | None = None
+        self._assert_pinned_models(chat_ok=False, embed_ok=False)
+        self._record_model_constraint(
+            "Zep Cloud memory path does not accept OpenAI chat/embed model pins; "
+            "stop before substituting — record scores only under Zep-native models."
+        )
+
+    def lane_id(self) -> str:
+        return "zep_cloud_free_tier"
+
+    def _client(self):
+        from zep_cloud import Zep
+
+        return Zep(api_key=self._api_key)
+
+    def _abort_if_billing(self, text: str) -> None:
+        if _BILLING_HINT.search(text or ""):
+            raise RuntimeError("Zep hard-stop: billing/upgrade/credit prompt detected")
+
+    def verify_free_tier(self) -> None:
+        client = self._client()
+        try:
+            client.user.list_ordered(page_size=1, page_number=1)
+        except Exception as exc:  # noqa: BLE001
+            self._abort_if_billing(str(exc))
+            raise
+
+    def _reset_remote(self) -> None:
+        client = self._client()
+        self.verify_free_tier()
+        self._user_id = self.namespace(("t1", "u1"))
+        self._thread_id = f"{self._user_id}-thread"
+        if self._thread_id:
+            try:
+                client.thread.delete(self._thread_id)
+            except Exception:
+                pass
+        try:
+            client.user.add(user_id=self._user_id, first_name="OpMem", last_name="Lane")
+        except Exception:
+            pass
+        client.thread.create(thread_id=self._thread_id, user_id=self._user_id)
+
+    def _remember_live(self, actor: tuple[str, str], content: str, mem_id: str) -> list[str]:
+        from zep_cloud.types import Message
+
+        client = self._client()
+        tid = self._thread_for(actor)
+        client.thread.add_messages(
+            thread_id=tid,
+            messages=[Message(content=content, role="user")],
+            return_context=False,
+        )
+        self.telemetry.log_http("remember", 200, "zep thread.add_messages")
+        if self.write_settle_seconds:
+            import time
+
+            time.sleep(self.write_settle_seconds)
+        return [mem_id]
+
+    def _thread_for(self, actor: tuple[str, str]) -> str:
+        if self._thread_id is None:
+            self._reset_remote()
+        return self._thread_id or self.namespace(actor)
+
+    def _recall_live(self, actor: tuple[str, str], query: str) -> list[dict]:
+        client = self._client()
+        uid = self._user_id or self.namespace(actor)
+        tid = self._thread_for(actor)
+        out: list[dict] = []
+        seen_content: set[str] = set()
+
+        msg_list = client.thread.get(tid, lastn=100)
+        for idx, msg in enumerate(msg_list.messages or []):
+            text = (msg.content or "").strip()
+            if not text or _overlap(query, text) <= 0:
+                continue
+            key = text.lower()
+            if key in seen_content:
+                continue
+            seen_content.add(key)
+            out.append({"id": f"zep-msg-{idx}", "content": text})
+
+        try:
+            results = client.graph.search(query=query, user_id=uid, limit=10)
+            self._abort_if_billing(str(results))
+            for ep in results.episodes or []:
+                text = getattr(ep, "content", None) or getattr(ep, "name", None) or ""
+                text = str(text).strip()
+                if not text or text.lower() in seen_content:
+                    continue
+                if _overlap(query, text) <= 0:
+                    continue
+                seen_content.add(text.lower())
+                eid = getattr(ep, "uuid_", None) or getattr(ep, "uuid", "zep-episode")
+                out.append({"id": str(eid), "content": text})
+            for edge in results.edges or []:
+                text = getattr(edge, "fact", None) or getattr(edge, "name", None) or ""
+                text = str(text).strip()
+                if not text or text.lower() in seen_content:
+                    continue
+                if _overlap(query, text) <= 0:
+                    continue
+                seen_content.add(text.lower())
+                eid = getattr(edge, "uuid_", None) or getattr(edge, "uuid", "zep-edge")
+                out.append({"id": str(eid), "content": text})
+        except Exception as exc:
+            self._abort_if_billing(str(exc))
+
+        out.sort(key=lambda row: -_overlap(query, row["content"]))
+        self.telemetry.log_http("recall", 200, str(out)[:500])
+        return out
+
+    def _forget_live(self, actor: tuple[str, str], memory_ids: list[str]) -> None:
+        self.telemetry.log_http("forget", 200, f"zep forget noop for {memory_ids}")
+
+    def _revise_live(self, actor: tuple[str, str], memory_ids: list[str], content: str) -> None:
+        from zep_cloud.types import Message
+
+        client = self._client()
+        tid = self._thread_for(actor)
+        client.thread.add_messages(
+            thread_id=tid,
+            messages=[Message(content=content, role="user")],
+            return_context=False,
+        )
+        self.telemetry.log_http("revise", 200, "zep thread.add_messages revise")
+        if self.write_settle_seconds:
+            import time
+
+            time.sleep(self.write_settle_seconds)
+
+    def _simulate_remember(self, content: str) -> None:
+        return
+
+    def _simulate_recall(self, query: str) -> None:
+        return
+
+    def _simulate_revise(self, content: str) -> None:
+        return
diff --git a/evals/run_opmem.py b/evals/run_opmem.py
index 86fefa59..cd857494 100644
--- a/evals/run_opmem.py
+++ b/evals/run_opmem.py
@@ -14,6 +14,7 @@
 
 import argparse
 import json
+import os
 import pathlib
 import sys
 
@@ -22,6 +23,11 @@
 
 from opmem_adapters import BrainyAdapter, Mem0OpAdapter, VerbatimBaseline
 
+try:
+    from opmem_external import LANE_ADAPTERS
+except ImportError:  # pragma: no cover
+    LANE_ADAPTERS = {}
+
 
 def resolve_actor(task: dict, name: str) -> tuple[str, str]:
     spec = task.get("actors", {}).get(name)
@@ -90,8 +96,13 @@ def build_adapters(names: list[str], base_url: str) -> list:
     registry = {
         "verbatim": lambda: VerbatimBaseline(),
         "brainy": lambda: BrainyAdapter(base_url),
+        "brainy-recall": lambda: BrainyAdapter(base_url, recall_lane="recall"),
+        "brainy-supersede": lambda: BrainyAdapter(base_url, revise_mode="supersede"),
         "mem0": lambda: Mem0OpAdapter(),
     }
+    live_external = os.environ.get("OPMEM_EXTERNAL_LIVE", "").strip() in ("1", "true", "yes")
+    for lane_name, factory in LANE_ADAPTERS.items():
+        registry[lane_name] = lambda f=factory, live=live_external: f(dry_run=not live)
     adapters = []
     for name in names:
         if name not in registry:
@@ -167,7 +178,9 @@ def main() -> int:
             out = ROOT.parent / out
         out.parent.mkdir(parents=True, exist_ok=True)
         out.write_text(output + "\n", encoding="utf-8")
-    return 1 if infrastructure_errors else 0
+    if infrastructure_errors and not os.environ.get("OPMEM_ALLOW_INFRA_EXIT"):
+        return 1
+    return 0
 
 
 if __name__ == "__main__":
diff --git a/evals/run_opmem_cost_dry_run.py b/evals/run_opmem_cost_dry_run.py
new file mode 100644
index 00000000..39a7e8fc
--- /dev/null
+++ b/evals/run_opmem_cost_dry_run.py
@@ -0,0 +1,87 @@
+#!/usr/bin/env python3
+"""Dry-run all OpMem external lanes: operation counts + token/cost worksheet (no vendor APIs)."""
+from __future__ import annotations
+
+import argparse
+import json
+import pathlib
+import sys
+
+ROOT = pathlib.Path(__file__).resolve().parent
+sys.path.insert(0, str(ROOT))
+
+from opmem_external import LANE_ADAPTERS  # noqa: E402
+from opmem_external.cost_budget import build_lane_budgets, build_worksheet_from_budgets  # noqa: E402
+from opmem_external.worksheet import build_worksheet, write_worksheet  # noqa: E402
+from run_opmem import run_task  # noqa: E402
+
+
+def main() -> int:
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--fixture-dir", default="fixtures/opmem")
+    parser.add_argument("--out-dir", default="docs/benchmarks/artifacts/opmem-external-lanes")
+    parser.add_argument("--run-id", default="opmem-dryrun-0001")
+    args = parser.parse_args()
+
+    fixture_dir = pathlib.Path(args.fixture_dir)
+    if not fixture_dir.is_absolute():
+        fixture_dir = ROOT.parent / fixture_dir
+    tasks = [json.loads(p.read_text(encoding="utf-8")) for p in sorted(fixture_dir.glob("*.json"))]
+    step_count = sum(len(t["steps"]) for t in tasks)
+
+    out_dir = pathlib.Path(args.out_dir)
+    if not out_dir.is_absolute():
+        out_dir = ROOT.parent / out_dir
+    out_dir.mkdir(parents=True, exist_ok=True)
+
+    lane_stats = []
+    per_lane_results = {}
+    for lane_name, factory in sorted(LANE_ADAPTERS.items()):
+        adapter = factory(dry_run=True, run_id=args.run_id)
+        results = []
+        infra = 0
+        for task in tasks:
+            try:
+                outcome = run_task(adapter, task)
+                results.append({"task": task["name"], "outcome": outcome})
+            except Exception as exc:  # noqa: BLE001
+                infra += 1
+                results.append({"task": task["name"], "error": f"{type(exc).__name__}: {exc}"})
+        per_lane_results[lane_name] = {
+            "infrastructure_errors": infra,
+            "results": results,
+        }
+        log_path = out_dir / f"raw-{lane_name}-dry-run.jsonl"
+        adapter.write_raw_log(log_path.with_suffix(".json"))
+        lane_stats.append(adapter.dry_run_stats())
+
+    worksheet = build_worksheet(lane_stats, task_count=len(tasks), step_count=step_count)
+    write_worksheet(out_dir / "cost-worksheet-dry-run.json", worksheet)
+    budgets = build_lane_budgets(tasks)
+    payload_ws = build_worksheet_from_budgets(
+        budgets, task_count=len(tasks), step_count=step_count,
+    )
+    (out_dir / "cost-worksheet-payload-budget.json").write_text(
+        json.dumps(payload_ws, indent=2) + "\n", encoding="utf-8",
+    )
+    (out_dir / "dry-run-results.json").write_text(
+        json.dumps(
+            {
+                "benchmark": "opmem-v0",
+                "mode": "dry_run",
+                "run_id": args.run_id,
+                "task_count": len(tasks),
+                "step_count": step_count,
+                "lanes": per_lane_results,
+            },
+            indent=2,
+        )
+        + "\n",
+        encoding="utf-8",
+    )
+    print(json.dumps(worksheet.to_dict(), indent=2))
+    return 0
+
+
+if __name__ == "__main__":
+    sys.exit(main())
diff --git a/evals/run_opmem_external_live.py b/evals/run_opmem_external_live.py
new file mode 100644
index 00000000..d15f85ed
--- /dev/null
+++ b/evals/run_opmem_external_live.py
@@ -0,0 +1,290 @@
+#!/usr/bin/env python3
+"""Phase 2: smoke or full OpMem external lane runs with preflight + artifacts."""
+
+from __future__ import annotations
+
+import argparse
+import json
+import os
+import platform
+import subprocess
+import sys
+import time
+from datetime import datetime, timezone
+from pathlib import Path
+
+ROOT = Path(__file__).resolve().parent
+sys.path.insert(0, str(ROOT))
+
+from opmem_external import LANE_ADAPTERS  # noqa: E402
+from opmem_external.cost_budget import (  # noqa: E402
+    build_lane_budgets,
+    build_worksheet_from_budgets,
+    load_tasks,
+)
+from opmem_external.pins import LANE_ENV_REQUIRED, SPEND_CEILING_USD  # noqa: E402
+from run_opmem import run_task  # noqa: E402
+
+DEFAULT_OUT = ROOT.parent / "docs/benchmarks/artifacts/opmem-external-lanes"
+SMOKE_TASK = "dup01_idempotent_remember.json"
+
+
+def _utc_now() -> str:
+    return datetime.now(timezone.utc).strftime("%Y-%m-%dT%H:%M:%SZ")
+
+
+def _package_versions() -> dict[str, str]:
+    names = (
+        "mem0ai",
+        "qdrant-client",
+        "zep-cloud",
+        "tiktoken",
+        "langmem",
+        "langgraph-checkpoint-postgres",
+        "openai",
+        "langchain-openai",
+    )
+    out: dict[str, str] = {}
+    for name in names:
+        try:
+            import importlib.metadata as md
+
+            out[name] = md.version(name)
+        except Exception:
+            out[name] = "not_installed"
+    return out
+
+
+def _preflight_budget(tasks: list[dict]) -> dict:
+    step_count = sum(len(t["steps"]) for t in tasks)
+    budgets = build_lane_budgets(tasks)
+    return build_worksheet_from_budgets(budgets, task_count=len(tasks), step_count=step_count)
+
+
+def _lane_blocked(lane: str) -> tuple[bool, str]:
+    if lane == "zep":
+        if not os.environ.get("ZEP_API_KEY", "").strip():
+            return True, "ZEP_API_KEY not present in runtime (vault not injected on this VM)"
+    if lane == "letta":
+        base = os.environ.get("LETTA_BASE_URL", "http://127.0.0.1:8283").rstrip("/")
+        local = base.startswith("http://127.0.0.1") or base.startswith("http://localhost")
+        if not local and not os.environ.get("LETTA_APP_SERVER_TOKEN", "").strip():
+            return True, "LETTA_APP_SERVER_TOKEN required for non-local Letta base URL"
+        try:
+            import urllib.request
+
+            urllib.request.urlopen(f"{base}/v1/health", timeout=3)
+        except Exception as exc:
+            return True, f"Letta App Server unreachable at {base}: {exc}"
+    if lane == "langmem":
+        if not os.environ.get("DATABASE_URL", "").strip():
+            default = "postgresql://opmem:opmem@127.0.0.1:5432/opmem_langmem"
+            os.environ.setdefault("DATABASE_URL", default)
+    for key in LANE_ENV_REQUIRED.get(lane, ()):
+        if not os.environ.get(key, "").strip():
+            return True, f"missing required env {key}"
+    return False, ""
+
+
+def _run_lane(
+    lane: str,
+    tasks: list[dict],
+    *,
+    run_id: str,
+    out_dir: Path,
+    mode: str,
+) -> dict:
+    blocked, reason = _lane_blocked(lane)
+    if blocked:
+        return {
+            "lane": lane,
+            "status": "blocked",
+            "reason": reason,
+            "infrastructure_errors": 0,
+            "task_results": [],
+        }
+
+    factory = LANE_ADAPTERS[lane]
+    adapter = factory(dry_run=False, run_id=run_id)
+    if not adapter.available():
+        return {
+            "lane": lane,
+            "status": "blocked",
+            "reason": f"credentials: {adapter.dry_run_stats().missing_credentials}",
+            "infrastructure_errors": 0,
+            "task_results": [],
+        }
+
+    if lane == "zep":
+        zep = adapter
+        try:
+            zep.verify_free_tier()
+        except Exception as exc:
+            return {
+                "lane": lane,
+                "status": "blocked",
+                "reason": f"Zep free-tier preflight failed: {exc}",
+                "infrastructure_errors": 1,
+                "task_results": [],
+            }
+
+    infra = 0
+    results = []
+    started = time.time()
+    for task in tasks:
+        try:
+            outcome = run_task(adapter, task)
+            results.append(
+                {
+                    "task": task["name"],
+                    "passed": outcome["passed"],
+                    "failures": outcome.get("failures", []),
+                }
+            )
+        except Exception as exc:
+            infra += 1
+            results.append(
+                {
+                    "task": task["name"],
+                    "error": f"{type(exc).__name__}: {exc}",
+                    "invalid_incomplete": True,
+                }
+            )
+            if lane == "zep" and "hard-stop" in str(exc).lower():
+                break
+
+    elapsed = round(time.time() - started, 2)
+    raw_path = out_dir / f"raw-{lane}-{mode}.json"
+    adapter.write_raw_log(raw_path)
+    telemetry = getattr(adapter, "telemetry", None)
+    if telemetry is not None:
+        telemetry.write(out_dir / f"telemetry-{lane}-{mode}.json")
+
+    passed = sum(1 for r in results if r.get("passed"))
+    return {
+        "lane": lane,
+        "status": "completed" if infra == 0 else "invalid_incomplete",
+        "mode": mode,
+        "run_id": run_id,
+        "elapsed_seconds": elapsed,
+        "infrastructure_errors": infra,
+        "tasks_run": len(results),
+        "tasks_passed": passed,
+        "task_results": results,
+        "product_constraints": adapter.dry_run_stats().product_constraints,
+    }
+
+
+def main() -> int:
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--fixture-dir", default="fixtures/opmem")
+    parser.add_argument("--out-dir", default=str(DEFAULT_OUT))
+    parser.add_argument("--run-id", default="")
+    parser.add_argument(
+        "--lanes",
+        default="mem0-oss,zep,letta,langmem",
+        help="Comma-separated lane ids",
+    )
+    parser.add_argument("--smoke", action="store_true", help="Run one smallest task per lane")
+    parser.add_argument("--full", action="store_true", help="Run all 13 tasks per lane")
+    parser.add_argument("--skip-preflight", action="store_true")
+    args = parser.parse_args()
+
+    if not args.smoke and not args.full:
+        parser.error("Specify --smoke and/or --full")
+
+    fixture_dir = Path(args.fixture_dir)
+    if not fixture_dir.is_absolute():
+        fixture_dir = ROOT.parent / fixture_dir
+    all_tasks = load_tasks(fixture_dir)
+    run_id = args.run_id or f"opmem-live-{datetime.now(timezone.utc).strftime('%Y%m%d%H%M%S')}"
+    os.environ["OPMEM_RUN_ID"] = run_id
+
+    out_dir = Path(args.out_dir)
+    if not out_dir.is_absolute():
+        out_dir = ROOT.parent / out_dir
+    out_dir.mkdir(parents=True, exist_ok=True)
+
+    if not args.skip_preflight:
+        budget = _preflight_budget(all_tasks)
+        (out_dir / "cost-worksheet-payload-budget.json").write_text(
+            json.dumps(budget, indent=2) + "\n", encoding="utf-8",
+        )
+        total = next(r for r in budget["lanes"] if r["lane"] == "_total_openai_worst_case")
+        if not total["within_ceiling"]:
+            print("ABORT: projected spend exceeds owner ceiling", file=sys.stderr)
+            print(json.dumps(budget, indent=2))
+            return 2
+        for row in budget["lanes"]:
+            lane = row.get("lane", "")
+            if lane in SPEND_CEILING_USD and not row.get("within_ceiling", True):
+                print(f"ABORT: lane {lane} exceeds ceiling", file=sys.stderr)
+                return 2
+
+    lanes = [x.strip() for x in args.lanes.split(",") if x.strip()]
+    manifest = {
+        "benchmark": "opmem-v0",
+        "generated_utc": _utc_now(),
+        "run_id": run_id,
+        "host": platform.platform(),
+        "python": platform.python_version(),
+        "package_versions": _package_versions(),
+        "pinned_models": {
+            "chat": "gpt-4.1-mini-2025-04-14",
+            "embed": "text-embedding-3-small",
+        },
+        "ceilings_usd": SPEND_CEILING_USD,
+        "modes": [],
+    }
+
+    smoke_tasks = [t for t in all_tasks if t["name"] == SMOKE_TASK.replace(".json", "")]
+    if not smoke_tasks:
+        smoke_path = fixture_dir / SMOKE_TASK
+        smoke_tasks = [json.loads(smoke_path.read_text(encoding="utf-8"))]
+
+    if args.smoke:
+        smoke_out = []
+        for lane in lanes:
+            smoke_out.append(_run_lane(lane, smoke_tasks, run_id=run_id, out_dir=out_dir, mode="smoke"))
+        manifest["modes"].append({"mode": "smoke", "task": smoke_tasks[0]["name"], "lanes": smoke_out})
+        (out_dir / f"live-smoke-{run_id}.json").write_text(
+            json.dumps(smoke_out, indent=2) + "\n", encoding="utf-8",
+        )
+
+    if args.full:
+        full_out = []
+        for lane in lanes:
+            prior = next(
+                (m for m in manifest.get("modes", []) if m.get("mode") == "smoke"),
+                None,
+            )
+            if prior:
+                lane_smoke = next((x for x in prior["lanes"] if x["lane"] == lane), None)
+                if lane_smoke and lane_smoke.get("status") != "completed":
+                    full_out.append(
+                        {
+                            "lane": lane,
+                            "status": "skipped",
+                            "reason": "smoke did not complete cleanly",
+                        }
+                    )
+                    continue
+            full_out.append(_run_lane(lane, all_tasks, run_id=run_id, out_dir=out_dir, mode="full"))
+        manifest["modes"].append({"mode": "full", "task_count": len(all_tasks), "lanes": full_out})
+        (out_dir / f"live-full-{run_id}.json").write_text(
+            json.dumps(full_out, indent=2) + "\n", encoding="utf-8",
+        )
+
+    (out_dir / f"run-manifest-{run_id}.json").write_text(
+        json.dumps(manifest, indent=2) + "\n", encoding="utf-8",
+    )
+    commands_log = out_dir / "commands.log"
+    with commands_log.open("a", encoding="utf-8") as fh:
+        fh.write(f"\n# {_utc_now()} run_id={run_id}\n")
+        fh.write(" ".join(subprocess.list2cmdline(sys.argv)) + "\n")
+    print(json.dumps(manifest, indent=2))
+    return 0
+
+
+if __name__ == "__main__":
+    sys.exit(main())
diff --git a/evals/run_opmem_paper_experiments.py b/evals/run_opmem_paper_experiments.py
new file mode 100644
index 00000000..ff162b12
--- /dev/null
+++ b/evals/run_opmem_paper_experiments.py
@@ -0,0 +1,201 @@
+#!/usr/bin/env python3
+"""OpMem manuscript experiments: manifests, lane ablations, stability repeats.
+
+Uses the same task runner as run_opmem.py. Intended to run against a live Brainy
+API or the httptest server spawned by go test (see scripts/run-opmem-manuscript.sh).
+"""
+from __future__ import annotations
+
+import argparse
+import json
+import pathlib
+import subprocess
+import sys
+import time
+from datetime import datetime, timezone
+
+ROOT = pathlib.Path(__file__).resolve().parent.parent
+sys.path.insert(0, str(ROOT / "evals"))
+
+from run_opmem import build_adapters, run_task  # noqa: E402
+
+
+def git_sha() -> str:
+    try:
+        return (
+            subprocess.check_output(["git", "rev-parse", "HEAD"], cwd=ROOT, text=True)
+            .strip()
+        )
+    except (subprocess.CalledProcessError, FileNotFoundError):
+        return "unknown"
+
+
+def load_tasks(fixture_dir: pathlib.Path) -> list[dict]:
+    return [
+        json.loads(path.read_text(encoding="utf-8"))
+        for path in sorted(fixture_dir.glob("*.json"))
+    ]
+
+
+def run_suite(
+    base_url: str,
+    system_names: list[str],
+    tasks: list[dict],
+) -> dict:
+    adapters = build_adapters(system_names, base_url)
+    results = []
+    infrastructure_errors = 0
+    for task in tasks:
+        entry = {"task": task["name"], "category": task["category"], "systems": {}}
+        for adapter in adapters:
+            if not adapter.available():
+                entry["systems"][adapter.name] = {"skipped": True, "reason": "not configured"}
+                continue
+            try:
+                entry["systems"][adapter.name] = run_task(adapter, task)
+            except Exception as exc:  # noqa: BLE001
+                infrastructure_errors += 1
+                entry["systems"][adapter.name] = {"error": f"{type(exc).__name__}: {exc}"}
+        results.append(entry)
+
+    summary: dict[str, dict] = {}
+    for adapter in adapters:
+        by_category: dict[str, dict[str, int]] = {}
+        passed_total = total = 0
+        for entry in results:
+            outcome = entry["systems"].get(adapter.name, {})
+            if outcome.get("skipped") or outcome.get("error"):
+                continue
+            bucket = by_category.setdefault(entry["category"], {"passed": 0, "total": 0})
+            bucket["total"] += 1
+            total += 1
+            if outcome.get("passed"):
+                bucket["passed"] += 1
+                passed_total += 1
+        summary[adapter.name] = {
+            "overall": f"{passed_total}/{total}",
+            "by_category": {cat: f"{v['passed']}/{v['total']}" for cat, v in sorted(by_category.items())},
+        }
+
+    return {
+        "benchmark": "opmem-v0",
+        "systems": [a.name for a in adapters],
+        "summary": summary,
+        "infrastructure_errors": infrastructure_errors,
+        "results": results,
+    }
+
+
+def stability_runs(
+    base_url: str,
+    tasks: list[dict],
+    repeats: int,
+) -> dict:
+    runs = []
+    for index in range(repeats):
+        report = run_suite(base_url, ["brainy"], tasks)
+        passed = report["summary"].get("brainy", {}).get("overall", "0/0")
+        runs.append({"repeat": index + 1, "overall": passed, "infrastructure_errors": report["infrastructure_errors"]})
+        time.sleep(0.05)
+    overalls = [r["overall"] for r in runs]
+    return {
+        "repeats": repeats,
+        "system": "brainy",
+        "recall_lane": "search",
+        "revise_mode": "correct",
+        "runs": runs,
+        "stable": len(set(overalls)) == 1,
+        "unique_scores": sorted(set(overalls)),
+    }
+
+
+def mechanism_view(report: dict) -> dict:
+    """Category-level pass rates as a lightweight mechanism breakdown."""
+    out: dict[str, dict[str, str]] = {}
+    for system, summ in report.get("summary", {}).items():
+        out[system] = summ.get("by_category", {})
+    out["_note"] = (
+        "verbatim is the rank-only baseline (no correct/suppress lifecycle); "
+        "correction/suppression/staleness categories map to revise/forget semantics."
+    )
+    return out
+
+
+def main() -> int:
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--base-url", default="http://127.0.0.1:8080")
+    parser.add_argument("--fixture-dir", default="fixtures/opmem")
+    parser.add_argument("--out-dir", required=True)
+    parser.add_argument("--stability-repeats", type=int, default=5)
+    parser.add_argument("--skip-mem0", action="store_true")
+    args = parser.parse_args()
+
+    fixture_dir = pathlib.Path(args.fixture_dir)
+    if not fixture_dir.is_absolute():
+        fixture_dir = ROOT / fixture_dir
+    tasks = load_tasks(fixture_dir)
+    out_dir = pathlib.Path(args.out_dir)
+    if not out_dir.is_absolute():
+        out_dir = ROOT / out_dir
+    out_dir.mkdir(parents=True, exist_ok=True)
+
+    sha = git_sha()
+    manifest = {
+        "benchmark": "opmem-v0",
+        "git_sha": sha,
+        "run_utc": datetime.now(timezone.utc).strftime("%Y-%m-%dT%H:%M:%SZ"),
+        "fixture_dir": str(fixture_dir.relative_to(ROOT) if fixture_dir.is_relative_to(ROOT) else fixture_dir),
+        "task_count": len(tasks),
+        "base_url": args.base_url,
+        "protocol": "docs/research/opmem/PUBLICATION_READINESS.md",
+        "scoring": "substring assertions on ranked recall results",
+    }
+    (out_dir / "run-manifest.json").write_text(json.dumps(manifest, indent=2) + "\n", encoding="utf-8")
+
+    stability = stability_runs(args.base_url, tasks, args.stability_repeats)
+    valid = [r for r in stability["runs"] if r.get("infrastructure_errors", 0) == 0]
+    stability["valid_runs"] = len(valid)
+    stability["valid_scores"] = sorted({r["overall"] for r in valid})
+    stability["interpretation"] = (
+        "Runs with infrastructure_errors>0 are invalid/incomplete; "
+        "partial denominators (e.g. 9/10) are not comparable to a clean 13/13 pin."
+    )
+    stability["manifest"] = manifest
+    (out_dir / "opmem-stability.json").write_text(json.dumps(stability, indent=2) + "\n", encoding="utf-8")
+
+    lanes = run_suite(
+        args.base_url,
+        ["brainy", "brainy-recall", "brainy-supersede"],
+        tasks,
+    )
+    lanes["manifest"] = manifest
+    (out_dir / "opmem-lane-ablation.json").write_text(json.dumps(lanes, indent=2) + "\n", encoding="utf-8")
+
+    primary_systems = ["verbatim", "brainy"]
+    if not args.skip_mem0:
+        primary_systems.append("mem0")
+    primary = run_suite(args.base_url, primary_systems, tasks)
+    primary["manifest"] = manifest
+    (out_dir / "opmem-primary.json").write_text(json.dumps(primary, indent=2) + "\n", encoding="utf-8")
+
+    mechanism = {
+        "manifest": manifest,
+        "by_category": mechanism_view(primary),
+        "verbatim_overall": primary["summary"].get("verbatim", {}).get("overall"),
+    }
+    (out_dir / "opmem-mechanism.json").write_text(json.dumps(mechanism, indent=2) + "\n", encoding="utf-8")
+
+    log_lines = [
+        f"git_sha={sha}",
+        f"tasks={len(tasks)}",
+        f"primary={json.dumps(primary['summary'])}",
+        f"lanes={json.dumps(lanes['summary'])}",
+        f"stability={json.dumps(stability)}",
+    ]
+    (out_dir / "run.log").write_text("\n".join(log_lines) + "\n", encoding="utf-8")
+
+    return 1 if primary["infrastructure_errors"] else 0
+
+
+if __name__ == "__main__":
+    sys.exit(main())
diff --git a/internal/api/eval_e2e_test.go b/internal/api/eval_e2e_test.go
index df6869df..062b744e 100644
--- a/internal/api/eval_e2e_test.go
+++ b/internal/api/eval_e2e_test.go
@@ -2,6 +2,7 @@ package api
 
 import (
 	"context"
+	"os"
 	"os/exec"
 	"path/filepath"
 	"strconv"
@@ -45,6 +46,30 @@ func TestOpMemBenchmarkAgainstHTTPServer(t *testing.T) {
 	)
 }
 
+func TestOpMemPaperExperiments(t *testing.T) {
+	outDir := os.Getenv("OPMEM_MANUSCRIPT_OUT")
+	if outDir == "" {
+		outDir = t.TempDir()
+	}
+	args := []string{"--out-dir", outDir}
+	if os.Getenv("MEM0_API_KEY") == "" {
+		args = append(args, "--skip-mem0")
+	}
+	runEvalHarness(t, "evals/run_opmem_paper_experiments.py", "", args...)
+}
+
+func TestOpMemLaneAblationOnly(t *testing.T) {
+	outDir := os.Getenv("OPMEM_LANE_OUT")
+	if outDir == "" {
+		t.Skip("set OPMEM_LANE_OUT to write lane ablation JSON")
+	}
+	t.Setenv("OPMEM_ALLOW_INFRA_EXIT", "1")
+	runEvalHarness(t, "evals/run_opmem.py", "",
+		"--systems", "brainy,brainy-recall,brainy-supersede",
+		"--json-out", filepath.Join(outDir, "opmem-lane-ablation-isolated.json"),
+	)
+}
+
 func runEvalHarness(t *testing.T, script, fixtureDir string, extraArgs ...string) {
 	t.Helper()
 	t.Setenv("LANG", "C")
diff --git a/scripts/generate-opmem-paper-tex.py b/scripts/generate-opmem-paper-tex.py
new file mode 100644
index 00000000..85d0b8f3
--- /dev/null
+++ b/scripts/generate-opmem-paper-tex.py
@@ -0,0 +1,165 @@
+#!/usr/bin/env python3
+"""Generate LaTeX fragments for OpMem manuscript from run artifacts."""
+from __future__ import annotations
+
+import json
+import pathlib
+import sys
+
+ROOT = pathlib.Path(__file__).resolve().parent.parent
+ART = ROOT / "docs/benchmarks/artifacts/opmem-manuscript-20533f2"
+OUT = ROOT / "docs/research/opmem/paper/generated"
+
+
+def load(name: str) -> dict:
+    return json.loads((ART / name).read_text(encoding="utf-8"))
+
+
+def write(name: str, body: str) -> None:
+    OUT.mkdir(parents=True, exist_ok=True)
+    (OUT / name).write_text(body.strip() + "\n", encoding="utf-8")
+
+
+def main() -> int:
+    if not ART.is_dir():
+        print(f"missing artifact dir {ART}", file=sys.stderr)
+        return 1
+    manifest = load("run-manifest.json")
+    primary = load("opmem-primary.json")
+    lane_path = ART / "opmem-lane-ablation-isolated.json"
+    if not lane_path.is_file():
+        lane_path = ART / "opmem-lane-ablation.json"
+    lanes = json.loads(lane_path.read_text(encoding="utf-8"))
+    stability = load("opmem-stability.json")
+    mechanism = load("opmem-mechanism.json")
+
+    sha = manifest["git_sha"]
+    brainy = primary["summary"]["brainy"]["overall"]
+    verbatim = primary["summary"]["verbatim"]["overall"]
+    mem0_line = ""
+    if "mem0" in primary["summary"]:
+        mem0_line = (
+            f"Mem0 Platform (same run): \\textbf{{{primary['summary']['mem0']['overall']}}} "
+            f"(API surface: Platform via \\texttt{{Mem0OpAdapter}})."
+        )
+
+    write(
+        "current-sha-results.tex",
+        f"""
+\\evidence{{completed}}{{
+At git SHA \\texttt{{{sha[:12]}}} (UTC {manifest['run_utc']}), embedded API harness:
+Brainy (search lane) \\textbf{{{brainy}}}, verbatim \\textbf{{{verbatim}}}.
+{mem0_line}
+}}
+""",
+    )
+
+    write(
+        "combined-results.tex",
+        f"""
+\\evidence{{completed}}{{
+Committed \\texttt{{opmem-primary.json}} lists per-task outcomes for
+{', '.join(primary['systems'])} with manifest fields (SHA, date, fixture count).
+}}
+""",
+    )
+
+    lane_rows = []
+    for system in lanes["systems"]:
+        summ = lanes["summary"].get(system, {})
+        lane_rows.append(f"{system} & {summ.get('overall', 'n/a')} \\\\")
+    write(
+        "lane-ablation.tex",
+        """
+\\begin{table}[h]
+\\centering
+\\caption{Brainy recall lane and revise mapping (SHA """
+        + sha[:12]
+        + """).}
+\\begin{tabular}{lc}
+\\toprule
+Configuration & Overall \\\\
+\\midrule
+"""
+        + "\n".join(lane_rows)
+        + """
+\\bottomrule
+\\end{tabular}
+\\end{table}
+Search (\\texttt{/memories/search}) is the normative pin lane.
+\\texttt{brainy-recall} uses \\texttt{POST /recall} (context mode).
+Correction tasks on the recall lane may return HTTP 409 during \\texttt{/correct}
+(infrastructure errors in artifact JSON).
+""",
+    )
+
+    sup = lanes["summary"].get("brainy-supersede", {}).get("overall", "n/a")
+    corr = lanes["summary"].get("brainy", {}).get("overall", "n/a")
+    write(
+        "supersede-ablation.tex",
+        f"""
+\\evidence{{completed}}{{
+Mapping \\texttt{{revise}} $\\rightarrow$ \\texttt{{POST .../supersede}} yields
+\\textbf{{{sup}}} vs \\textbf{{{corr}}} with \\texttt{{/correct}} on the search lane
+(same SHA). Failures are task-level assertion misses, not missing runs.
+}}
+""",
+    )
+
+    by_cat = mechanism.get("by_category", {})
+    brainy_cat = by_cat.get("brainy", {})
+    verb_cat = by_cat.get("verbatim", {})
+    lines = [
+        "\\begin{tabular}{lcc}",
+        "\\toprule",
+        "Category & Brainy & Verbatim (rank-only) \\\\",
+        "\\midrule",
+    ]
+    for cat in sorted(brainy_cat.keys()):
+        if cat.startswith("_"):
+            continue
+        lines.append(f"{cat} & {brainy_cat.get(cat, '---')} & {verb_cat.get(cat, '---')} \\\\")
+    lines.append("\\bottomrule")
+    lines.append("\\end{tabular}")
+    write(
+        "mechanism-ablation.tex",
+        """
+Category pass rates separate correction/suppression/staleness contracts;
+verbatim approximates rank-only recall without governed lifecycle ops.
+\\begin{table}[h]
+\\centering
+"""
+        + "\n".join(lines)
+        + "\n\\end{table}",
+    )
+
+    runs = stability.get("runs", [])
+    valid = [r for r in runs if r.get("infrastructure_errors", 0) == 0]
+    invalid = [r for r in runs if r.get("infrastructure_errors", 0) > 0]
+    valid_scores = sorted({r["overall"] for r in valid})
+    run_lines = [
+        f"Repeat {r['repeat']}: {r['overall']} (infra errors={r['infrastructure_errors']})"
+        for r in runs
+    ]
+    write(
+        "stability.tex",
+        f"""
+\\evidence{{completed}}{{
+Five consecutive Brainy search-lane repeats on one embedded server session.
+\\textbf{{Valid runs only}} (infra errors $=0$): {len(valid)}/{len(runs)}; scores: {', '.join(valid_scores) or 'none'}.
+}}
+\\begin{{itemize}}
+"""
+        + "\n".join(f"\\item {line}" for line in run_lines)
+        + f"""
+\\end{{itemize}}
+Repeats with infra errors $>0$ ({len(invalid)} of {len(runs)}) are \\textbf{{invalid/incomplete}}: the harness omits errored tasks from the denominator, so aggregates such as 9/10 are \\emph{{not}} comparable to a clean 13/13 pin.
+Use isolated \\texttt{{TestOpMemBenchmarkAgainstHTTPServer}} for merge-gate stability.
+""",
+    )
+
+    return 0
+
+
+if __name__ == "__main__":
+    sys.exit(main())
diff --git a/scripts/opmem-start-letta-server.sh b/scripts/opmem-start-letta-server.sh
new file mode 100755
index 00000000..6929e538
--- /dev/null
+++ b/scripts/opmem-start-letta-server.sh
@@ -0,0 +1,34 @@
+#!/usr/bin/env bash
+# Start pinned Letta REST server for OpMem external lane (no Docker).
+set -euo pipefail
+export PATH="${HOME}/.local/bin:${PATH}"
+LETTA_VERSION="${OPMEM_LETTA_SERVER_VERSION:-0.11.7}"
+LETTA_CLIENT_VERSION="${OPMEM_LETTA_CLIENT_VERSION:-0.1.307}"
+PORT="${LETTA_PORT:-8283}"
+HOST="${LETTA_HOST:-127.0.0.1}"
+LOG="${OPMEM_LETTA_SERVER_LOG:-/tmp/opmem-letta-server.log}"
+
+python3 -m pip install -q "letta==${LETTA_VERSION}" "letta-client==${LETTA_CLIENT_VERSION}" sqlite-vec
+
+if ! command -v letta >/dev/null; then
+  echo "letta CLI not on PATH; ensure ~/.local/bin is exported" >&2
+  exit 1
+fi
+
+if curl -sf "http://${HOST}:${PORT}/v1/health" >/dev/null 2>&1; then
+  echo "Letta already listening on http://${HOST}:${PORT}"
+  exit 0
+fi
+
+nohup letta server --type rest --port "${PORT}" --host "${HOST}" --no-secure >"${LOG}" 2>&1 &
+echo "Started Letta ${LETTA_VERSION} at http://${HOST}:${PORT} (log: ${LOG})"
+
+for _ in $(seq 1 30); do
+  if curl -sfL "http://${HOST}:${PORT}/v1/health" | grep -q ok; then
+    echo "Letta health OK"
+    exit 0
+  fi
+  sleep 1
+done
+echo "Letta failed to become healthy; see ${LOG}" >&2
+exit 1
diff --git a/scripts/opmem-verify-zep-env.sh b/scripts/opmem-verify-zep-env.sh
new file mode 100755
index 00000000..d468450c
--- /dev/null
+++ b/scripts/opmem-verify-zep-env.sh
@@ -0,0 +1,15 @@
+#!/usr/bin/env bash
+# Verify Zep lane prerequisites without printing secrets.
+set -euo pipefail
+echo "CLOUD_AGENT_INJECTED_SECRET_NAMES:"
+echo "${CLOUD_AGENT_INJECTED_SECRET_NAMES:-}" | tr ',' '\n' | sort
+if [ -n "${ZEP_API_KEY:-}" ]; then
+  echo "ZEP_API_KEY: present (length=${#ZEP_API_KEY})"
+else
+  echo "ZEP_API_KEY: MISSING"
+  exit 1
+fi
+if ! echo "${CLOUD_AGENT_INJECTED_SECRET_NAMES:-}" | tr ',' '\n' | grep -qx 'ZEP_API_KEY'; then
+  echo "WARNING: ZEP_API_KEY not listed in CLOUD_AGENT_INJECTED_SECRET_NAMES"
+fi
+exit 0
diff --git a/scripts/run-opmem-manuscript.sh b/scripts/run-opmem-manuscript.sh
new file mode 100755
index 00000000..4380429f
--- /dev/null
+++ b/scripts/run-opmem-manuscript.sh
@@ -0,0 +1,28 @@
+#!/usr/bin/env bash
+# Run OpMem manuscript experiments via embedded API (same stack as CI).
+set -euo pipefail
+ROOT="$(cd "$(dirname "$0")/.." && pwd)"
+cd "$ROOT"
+SHA="$(git rev-parse HEAD)"
+OUT_DIR="${1:-docs/benchmarks/artifacts/opmem-manuscript-${SHA:0:8}}"
+mkdir -p "$OUT_DIR"
+
+{
+  echo "git_sha=$SHA"
+  echo "date_utc=$(date -u +%Y-%m-%dT%H:%M:%SZ)"
+  echo "go test ./internal/api/ -run TestOpMemBenchmarkAgainstHTTPServer -count=1"
+  echo "OPMEM_MANUSCRIPT_OUT=$OUT_DIR go test ./internal/api/ -run TestOpMemPaperExperiments -count=1"
+} > "$OUT_DIR/commands.log"
+
+go test ./internal/api/ -run TestOpMemBenchmarkAgainstHTTPServer -count=1 2>&1 | tee "$OUT_DIR/go-test-opmem.log"
+
+export OPMEM_MANUSCRIPT_OUT="$OUT_DIR"
+go test ./internal/api/ -run TestOpMemPaperExperiments -count=1 2>&1 | tee "$OUT_DIR/go-test-paper.log"
+
+# Mem0 Platform counter-run when MEM0_API_KEY is set (optional; not required for CI).
+if [[ -n "${MEM0_API_KEY:-}" ]]; then
+  echo "mem0 run: python3 evals/run_opmem.py --systems mem0 --json-out $OUT_DIR/opmem-mem0-rerun.json" >> "$OUT_DIR/commands.log"
+  python3 evals/run_opmem.py --systems mem0 --json-out "$OUT_DIR/opmem-mem0-rerun.json" 2>&1 | tee "$OUT_DIR/mem0-rerun.log" || true
+fi
+
+echo "Artifacts written to $OUT_DIR"