agent-wiki

codingfeng-fufu

Agent-native LLM Wiki for persistent, evidence-backed project memory.

Open codingfeng-fufu/agent-wiki → On the boundary of the four rules.

Works with any agent

ships AGENTS.md, the cross-tool instruction file.

In its own words

Aboutcodingfeng-fufu/agent-wiki

It turns raw sources into a maintained Markdown knowledge base that agents can search, query, update safely, and access through MCP.

README.md

Local, source-backed Markdown memory for coding agents.

LLM Wiki is a runnable implementation of Andrej Karpathy's LLM Wiki pattern: keep raw sources as immutable evidence, let an agent maintain a connected Markdown wiki in front of them, and use that wiki as the durable knowledge layer for future work.

This is not a hosted knowledge base and not another query-time RAG wrapper. It is a local-first developer tool for agents that need memory across sessions: source pages, concept pages, links, provenance, search, MCP tools, safe plans, dry-run apply, and health checks.

It is an independent project, not an official project from Andrej Karpathy, OpenAI, Anthropic, the Model Context Protocol project, qmd, Obsidian, or any other referenced tool. See Origin.

Why Not Just RAG?

Query-time RAG retrieves raw chunks every time a user asks a question. That is useful, but it does not naturally accumulate structure.

LLM Wiki uses a compile-first workflow:

The agent reads sources once, writes durable Markdown pages, connects related concepts, records provenance, and reuses the maintained wiki through CLI or MCP tools. New questions can also become saved wiki pages, so the knowledge base improves instead of disappearing into chat history.

Read more in Why Not Just RAG? and Karpathy Pattern Compatibility.

Quick Start

This first public version is optimized for local use from a GitHub clone.

Check the included example wiki:

The repository's own description, then 26% of its README. Read the rest on GitHub ↗

Read from the repository's git tree on 2026-08-17.