Skip to content

README.md

Latest commit

 

History

History
68 lines (47 loc) · 2.96 KB

README.md

File metadata and controls

68 lines (47 loc) · 2.96 KB
Raw

memory-mini

Minimal reference for durable agent memory: namespaced key-value storage with upsert, soft-delete, cleanup, optional embeddings, and SQLite durability.

memory-mini is intentionally small. It is a local reference implementation for inspecting namespaced SQLite storage, lifecycle operations, upserts, retention, and optional embedding helpers. Its scope is the code and tests in this repository, not a hosted memory service or a production-readiness claim.

What to inspect

  • store.py defines the SQLite schema, the (namespace, key) uniqueness rule, and the ON CONFLICT upsert path.
  • store.py keeps soft-delete and retention-based cleanup as separate operations.
  • namespace.py contains the path-like namespace normalization and traversal helpers.
  • embeddings.py contains the optional vector helpers; the package has no required runtime dependencies.

Install

Install directly from the Git repository. Use the base command or the command with the optional embedding dependency:

# Base package
python -m pip install "memory-mini @ git+https://github.com/JustinJLeopard/memory-mini.git@main"

# Base package plus NumPy-backed embedding helpers
python -m pip install "memory-mini[embed] @ git+https://github.com/JustinJLeopard/memory-mini.git@main"

There is currently no PyPI release for this package.

30-second example

from datetime import timedelta
from memory_mini import Store

with Store("memory.db") as memory:
    memory.store("session-end-2026-05-01-1", "Shipped docs and tests.", "sessions/project")
    memory.store("session-end-2026-05-01-1", "Shipped docs, tests, and CI.", "sessions/project")

    print(memory.get("session-end-2026-05-01-1", "sessions/project").value)

    memory.soft_delete("session-end-2026-04-01-1", "sessions/project")
    memory.cleanup(retention=timedelta(days=30))

Why upsert-first?

Durable memory should treat repeated writes as a normal update path. A delete-then-store workflow creates avoidable failure modes: unique keys can remain reserved by soft-deleted rows, audit history becomes ambiguous, and callers have to coordinate two operations instead of one. memory-mini defaults to upsert so storing the same namespaced key again refreshes the active entry in a single transaction.

Soft-delete and cleanup are deliberately separate. Soft-delete hides an entry from normal reads while preserving recoverability. Cleanup is the explicit hard-removal step, usually run after a retention window.

Docs

Related

  • JustAi — orchestration control plane.
  • safe-mini — safe local execution substrate.
  • route-mini — routing policy reference.