1.9 KiB
LLM Wiki: A New Approach to Knowledge Management
Source: Andrej Karpathy's blog / GitHub Gist (2026)
Summary
Andrej Karpathy introduced the concept of an "LLM Wiki" — a persistent, structured knowledge base built and maintained by an LLM agent. Unlike traditional RAG (Retrieval-Augmented Generation) systems that search raw documents at query time, an LLM Wiki is compiled once and kept current.
Key Concepts
The Problem with RAG
Traditional RAG retrieves from raw documents every time you ask a question. Results depend heavily on chunking strategy, embedding quality, and retrieval parameters. Good results are not retained.
The LLM Wiki Approach
Instead of retrieving at query time, the LLM reads sources once and builds a structured wiki:
- Raw sources are stored in
raw/(immutable) - Compiled knowledge lives in
wiki/as interlinked markdown pages - A schema file (
AGENTS.md) defines conventions and workflows
Three Workflows
- Ingest: Add source to
raw/, LLM compiles into wiki (~10-15 pages per source) - Query: LLM reads
index.md, finds relevant pages, synthesizes answer with citations - Lint: Periodic audit for contradictions, orphans, stale claims
Compounding Knowledge
Good answers get filed back into the wiki as new pages. The wiki gets richer with every source added and every question asked. This creates a compounding knowledge effect.
Technical Implementation
- Works with any LLM that can read/write files
- Tested with Claude Code, Codex CLI, Cursor
- Also works with local models via Ollama (Gemma, Llama, etc.)
- No special infrastructure needed — just a git repository of markdown files
- Optional: qmd for enhanced search at scale (100+ sources)
Key Advantages
- No embedding models or vector databases needed
- Full version control via git
- Human-in-the-loop curation
- Knowledge compounds over time
- Works offline / on-premise