Summary

A pattern where LLMs incrementally build and maintain a structured, interlinked knowledge base (wiki) instead of relying on RAG for every query. Knowledge is “compiled” once from raw sources, then kept current — the LLM handles the bookkeeping humans tend to abandon.

Key Points

  • Three layers: Raw Sources (immutable) → Wiki (LLM-maintained) → Schema (conventions)
  • Three operations: Ingest (add knowledge), Query (retrieve + synthesize), Lint/Maintain (audit quality)
  • Compiled truth pattern: Separate current conclusions from evidence timeline
  • LLM’s sweet spot: Cross-referencing, consistency, multi-file updates — tedious work humans skip
  • Schema as leverage: A well-written config file (CLAUDE.md) can drive the entire system without code

Implementations

  • Karpathy (2025): Conceptual pattern, pure Markdown + CLAUDE.md
  • Garry Tan / GBrain (2025): Full product spec — SQLite, FTS5, vectors, MCP server, CLI

Open Questions

  • How to prevent error accumulation over many update cycles?
  • At what scale does Markdown + grep stop being sufficient?
  • Does LLM-maintained knowledge reduce the human’s own understanding?

Evidence Timeline

  • 2026-04-07: Initial compilation from Karpathy’s LLM Wiki gist and Garry Tan’s GBrain spec

相关页面

harness, andrej-karpathy