LLM Wiki Pattern (Karpathy)

A method for building persistent, structured knowledge bases using LLM agents — compiled once, kept current, queried efficiently.

Overview

Andrej Karpathy proposed the "LLM Wiki" pattern as an alternative to RAG (Retrieval-Augmented Generation). Instead of searching raw documents on every query, an LLM agent incrementally builds and maintains a wiki of interlinked markdown pages.

Core Architecture

Three Layers

Raw Sources (raw/) — Original documents, immutable. Never modified by the LLM.
Wiki (wiki/) — Structured, interlinked markdown pages compiled by the LLM.
Schema (AGENTS.md) — Rules, conventions, and workflows. The "operating system" of the wiki.

Workflows

Ingest: Drop source in raw/ → LLM reads, extracts, creates/updates ~10-15 wiki pages per source
Query: Ask question → LLM reads index.md → drills into relevant pages → synthesizes answer with citations
Lint: Periodically audit wiki for contradictions, orphans, stale claims

Key Principles

One concept = one file
Cross-reference everything with [[wikilinks]]
The LLM owns the wiki directory; humans only edit raw/ and the schema
Good answers get filed back into the wiki as new pages (compounding knowledge)
Works with any LLM that can read/write files (Claude Code, Codex, Cursor, local models via Ollama)

AGENTS.md — Schema file for this wiki
index — Master catalog

Sources

Karpathy's original idea file (GitHub Gist)

1.6 KiB Raw Blame History