Karpathy's LLM Wiki¶
A pattern popularized by Andrej Karpathy (co-founder of OpenAI, ex-Tesla AI lead): instead of traditional rag-vs-wiki, have an LLM compile knowledge once into a persistent, interlinked set of markdown files — a wiki — and keep it current. This very knowledge base is built on that pattern.
The core idea¶
Traditional RAG rediscovers knowledge from scratch on every query: retrieve chunks, stuff into
context, hope relevance. A compiled wiki does the work up front — cross-references are
already there, contradictions already flagged, synthesis already reflects everything ingested.
The agent reads a small index.md, follows wiki-links, and reasons over curated pages.
Division of labor¶
- Human curates sources and directs analysis.
- Agent summarizes, cross-references, files, and maintains consistency.
Key facts¶
- Structure:
raw/(immutable sources) → curatedentities/,concepts/,comparisons/pages →index.md+log.mdnavigation. - No vector database required — retrieval is agentic (search + read + follow links), see rag-vs-wiki.
- Just markdown: opens in Obsidian, VS Code, or any editor; portable and owned.
- Compounds over time — each ingest makes the whole more valuable, unlike per-query RAG.
Related¶
- rag-vs-wiki — the tradeoff this pattern takes a side on
- hermes — the agent maintaining this wiki
- stack-sovereignty-box — where this wiki lives and is served