Skip to content

Add self-evolving knowledge base (Karpathy LLM-wiki pattern)#50

Draft
isaacsight wants to merge 2 commits into
mainfrom
claude/claude-knowledge-base-setup-tray2w
Draft

Add self-evolving knowledge base (Karpathy LLM-wiki pattern)#50
isaacsight wants to merge 2 commits into
mainfrom
claude/claude-knowledge-base-setup-tray2w

Conversation

@isaacsight

Copy link
Copy Markdown
Owner

A local-first, plain-text second brain under knowledge-base/, built on Andrej Karpathy's LLM-wiki pattern. No Obsidian, no vector store, no RAG — just folders and markdown that get more useful over time because the agent organizes and audits them on every pass.

The five building blocks

Block Path Role
Inputs knowledge-base/raw/ Dumping ground — drop files/notes/links, messy by design
Wiki knowledge-base/wiki/ Organized layer: index.md (the map), log.md (append-only journal), processed.md (ingestion registry)
Outputs knowledge-base/outputs/ Generated artifacts (PDFs, decks, briefs)
Operating manual knowledge-base/CLAUDE.md Agent-maintained system prompt, auto-loaded when working in-folder
Dream sequence .claude/commands/dream-sequence.md /dream-sequence lint pass

How it works

  • Feed it — drop into raw/, or say "save this" to a link and Claude captures it with source + date.
  • Ask it — Claude reads wiki/index.md to find where an answer lives, then pulls the relevant files.
  • Audit it/dream-sequence ingests new raw captures, lints the wiki for contradictions, stale claims, duplicates, and orphans, and rewrites the index/log/registry. The processed.md registry ensures a raw file is never folded in twice.

The wiki ships empty and is seeded by feeding raw/ then running the dream sequence. Ingestion behavior and cadence are documented and user-configurable — and the manual edits itself when the rules change.

Scheduling

/dream-sequence runs on demand; point a Claude desktop routine (or any scheduler) at it for daily/weekly auto-audits.

https://claude.ai/code/session_013ntEVS8fnPoxyhwJqpQgLo


Generated by Claude Code

claude added 2 commits June 10, 2026 19:45
A local-first, plain-text second brain under knowledge-base/. Five
building blocks, no external tooling:

- raw/      inputs dumping ground (messy by design)
- wiki/     organized layer: index (map), log (journal), processed
            (ingestion registry so raw files are never folded twice)
- outputs/  generated artifacts
- CLAUDE.md agent-maintained operating manual; auto-loaded in-folder
- /dream-sequence  lint pass: ingest new raw, flag contradictions /
            stale claims / duplicates / orphans, rewrite index+log+registry

The wiki ships empty and is seeded by feeding raw/ then running the
dream sequence. Ingestion behavior and cadence are documented and
user-configurable; the manual edits itself when the rules change.

https://claude.ai/code/session_013ntEVS8fnPoxyhwJqpQgLo
End-to-end demonstration of the second brain, mirroring the workflow:
feed raw, run the dream sequence, get outputs.

- raw/: 3 source-cited research captures (market & adoption,
  technical stack/MCP/orchestration, risks & governance)
- wiki/: synthesized topic page agentic-ai.md; index rebuilt with 6
  concepts / 11 entities / 3 sources; log + processed registry updated
- Dream-sequence lint: no contradictions/orphans/stale; merged the
  one duplicate theme (adoption outpacing safe deployment) into the
  topic page's connective thread
- outputs/: agentic-ai-brief-2026.md generated from the wiki

https://claude.ai/code/session_013ntEVS8fnPoxyhwJqpQgLo
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants