Andrej Karpathy · 2026-04-04 · major
Andrej Karpathy's LLM Wiki — the 'drop RAG, let the agent maintain a markdown wiki' pattern
Karpathy's idea-file gist describes replacing RAG over raw docs with an LLM that incrementally compiles sources into a structured, cross-linked markdown wiki — 16M X views, 5k gist stars, dozens of community implementations within two weeks.
Stop treating LLMs as retrieval-over-raw-docs. Point an agent at a folder of sources and let it build and maintain a living wiki instead.
What is it?
A GitHub gist from Andrej Karpathy, posted April 3–4, 2026, that sketches a pattern for personal knowledge bases. Rather than classic RAG — chunk your docs, embed them, retrieve the best chunks at query time — an LLM agent incrementally compiles raw sources into a structured markdown wiki (entity pages, concept pages, cross-references, index files, contradiction flags). You curate sources and questions; the agent does the organizational labor. The X post announcing it hit 16 million views and 100k bookmarks in 48 hours; the gist passed 5,000 stars.
How does it work?
The gist is deliberately code-free — 'an idea file designed to be copy-pasted to your own LLM Agent' like Claude Code, Codex, OpenCode, or Cursor. It describes three components (raw sources, the wiki, a schema) and three operations (Ingest new sources into the right wiki pages, Query the wiki to answer questions, Lint to catch contradictions and gaps). The wiki itself is just a folder of markdown files, ideally viewed in Obsidian, which the agent keeps cross-linked and tidy as you add more sources.
Why does it matter?
This is the closest thing the community has had to a new consensus pattern for 'long-lived personal context' since classic RAG landed. Within days, community reimplementations (llmwiki.app, Obsidian plugins, karpathy-llm-wiki skills, LLM Wiki v2 with agent memory) started shipping. Karpathy's framing — 'in the age of LLM agents, sharing an idea is more valuable than sharing code' — is also becoming its own meme about how agent-era ideas propagate.
Who is it for?
Anyone who has tried and failed to keep a second-brain system populated, or who runs repeated research on a narrow topic.
Try it
Clone the gist into your Claude Code / Codex / OpenCode project root and follow the ingest/query/lint instructions