llm-wiki

A local-first memory system for AI agents. Markdown files, git-tracked, and owned by you.

LLM Wiki gives every AI coding agent you use — OpenCode, Codex, Claude Code, Cursor, Antigravity — one MCP-first interface to a shared, persistent knowledge base. MCP handles reads and actions; thin native lifecycle adapters capture session events that MCP cannot observe. Durable knowledge survives across sessions so you never re-explain the same thing twice.

Everything lives on your disk as plain markdown: readable in Obsidian, diffable in git, owned entirely by you.

Storage, capture, MCP, and retrieval are local. Model-backed classification and compilation use the configured provider: OpenCode, Codex, Claude, and OpenAI may use cloud services; Ollama may be local. Auto-detection is not a local-only guarantee.

Table of Contents

• How it works
• Features
• Quick Start
• Wire up your agents
• Architecture
• Evidence generations and migration
• Benchmark
• Comparison
• Contributing
• Credits
• License

How it works

The system follows the "compile, not retrieve" pattern (Karpathy, April 2026): raw session signals are captured in real time, then a background LLM pass compiles them into structured knowledge pages rather than relying on raw retrieval at query time.