What it is
The LLM Wiki is a pattern from Andrej Karpathy: instead of dumping documents into a search index, you keep a wiki that an LLM incrementally builds and maintains. When a new source arrives, the agent reads it, extracts what matters, integrates it into existing pages, updates cross-references, and flags contradictions. The knowledge accumulates in the wiki — not in a transcript that disappears.
Why it beats plain RAG
Classic RAG retrieves chunks at query time and regenerates an answer from scratch. Every question starts cold; nothing compounds. The maintained wiki is the opposite: the synthesis already happened, the cross-links are already there, and each new source makes the whole thing denser. RAG still runs — but over a curated wiki, not raw documents.
Three layers
- Raw — source documents. Human-owned, immutable. The LLM reads, never edits.
- Wiki — the LLM writes it; the human reads it. One concept per page, heavy cross-linking.
- Schema — a
CLAUDE.mdthat defines how the vault works. Co-owned, evolves with use.
The three workflows
Ingest a source → summarise, create entity/concept pages, update what it contradicts. Query → answer with citations, then file the answer back so explorations compound. Lint → periodic health check for contradictions, stale claims, orphan pages, missing links.
How I run mine
Mine lives in Obsidian. Smart Connections gives semantic search and chat over the wiki; Dataview drives dynamic index tables from page frontmatter. A projects/ folder bridges the vault to my actual work folders — each project page links straight to its directory, so I open the vault, find the project, and click through. The vault is the single entry point for both knowledge and work.