Le wiki LLM de Karpathy
A way to give an AI agent durable memory with no database and no server: three folders of Markdown files you own. `raw/` keeps the source material you never touch, `wiki/` holds the notes the AI writes itself by summarizing those sources, and an index file that lists everything and fits in a single context window. Andrej Karpathy published it in April 2026; the post got 16 million views for a plain folder structure.
Strengths
- Zero infrastructure: three Markdown folders, no database or server, files you own and can version with git
- Memory compounds instead of clearing: the AI re-reads a compact index and finds the exact note, without re-reading everything each time
- The AI keeps the wiki up to date (canonizes, merges, drops the stale), not you by hand
- Human-readable: you can open any note, fix it, understand why the AI decided something
Limitations
- Lives or dies on your discipline: without a sharp doctrine of what deserves a note, the wiki inflates into noise and becomes unusable
- Not built for search over millions of documents: past a certain volume, a vector database stays more suitable
- The index must fit in one context window: past a few hundred notes, you have to split it, otherwise it overflows
Best for
- Anyone who works continuously with an AI on their projects and is tired of re-explaining everything each session
- PMs, POs and designers who want a shared team memory, readable and versioned, without standing up infrastructure