WTF is LLM Wiki?
Instead of retrieving from raw documents, an agent builds and maintains its own structured wiki of what it has learned.
Also called: agent wiki, idea file, agent-maintained knowledge base
TL;DR
RAG searches your pile of documents fresh every time. An LLM wiki flips it: as the agent reads sources, it writes and updates tidy Markdown pages, links them, merges duplicates and keeps summaries current. Knowledge compounds instead of being re-derived. Karpathy's framing: the LLM is the programmer, the wiki is the codebase.
How it works
Published as an 'idea file': a plain-text description of the pattern that you paste into an agent so it builds its own implementation. Core operations are ingest (read a source, update affected pages), query (answer from the wiki, then file the answer back into it) and lint (find contradictions, orphans and stale pages). Pages are human-readable Markdown, often in Obsidian, with the agent doing the bookkeeping. Compared with vector RAG it trades retrieval breadth for curated, consistent, inspectable knowledge.
Origin
- Coined / introduced by
- Andrej Karpathy
- When
- 3 Apr 2026
- Story
- A GitHub gist with no code that drew millions of views and hundreds of community implementations within days.