AP0110.ORG
All Wiki
analysis

LLM Wiki: A Pattern for Building Personal Knowledge Bases Using LLMs

Karpathy's pattern for an LLM-maintained wiki: instead of re-deriving answers from raw documents on every query (RAG), the model integrates each source into a persistent, compounding knowledge base.

Links to 5 knowledge-managementllmagents
DocsWiki
On this page & linked pages

Summary

Published as a public gist by Andrej Karpathy, this short essay describes a pattern for personal knowledge bases built and maintained by a language model. It is positioned against retrieval-augmented generation (RAG): where RAG rediscovers knowledge from raw chunks on every query and accumulates nothing, an LLM-maintained wiki integrates each new source once into a persistent, interlinked set of Markdown pages — updating entity and concept pages, revising summaries, flagging contradictions. “The knowledge is compiled once and kept current, not re-derived on every query.”

Key ideas

  • Three layers — immutable raw sources (human-curated), the wiki (LLM-owned Markdown), and the schema (a CLAUDE.md / AGENTS.md-style config of conventions and workflows).
  • Three operations — ingest, query, and lint.
  • Two indexes — a content-oriented index.md and a chronological log.md.
  • The maintenance bottleneck that defeats human wikis is bookkeeping, which an LLM does for near-zero cost.

Why it matters

The essay closes by framing LLM-maintained wikis as the delivery of Vannevar Bush’s 1945 Memex vision — “the part he couldn’t solve was who does the maintenance; the LLM handles that.” It is the canonical statement of the LLM-Maintained Wiki pattern, a cousin of Interpretable Context Methodology and the agent-readable web standards — all file-native, plain-text, Markdown-first. This very wiki is built on the pattern. Only an original summary and the canonical link are kept here.

Part of the AP0110.ORG source library. See the wiki overview.

Full source text: read the original document.