Summary

This project automatically builds an evolving memory for AI coding agents like Claude Code. It captures conversations, extracts key insights via the Claude Agent SDK, and compiles them into a structured, cross-referenced knowledge base. Inspired by Karpathy's LLM Knowledge Base, it provides users with a dynamic personal knowledge repository from AI interactions, bypassing complex RAG setups for personal-scale knowledge management.

How It Works

Hooks capture conversation transcripts at session end or pre-compaction. flush.py uses the Claude Agent SDK to extract decisions and lessons learned, logging them daily. compile.py processes these logs into structured, cross-referenced knowledge articles organized by concept. Retrieval uses query.py with a simple index file, avoiding vector databases for personal-scale knowledge bases where LLMs reportedly outperform RAG. A session start hook injects this index into future AI sessions, creating a continuous learning cycle.

Quick Start & Requirements

Instruct your Claude Code agent to clone the repository and configure hooks. The agent will clone https://github.com/coleam00/claude-memory-compiler, run uv sync for dependencies, and integrate .claude/settings.json hooks. Requires an active Claude subscription (Max, Team, or Enterprise) for personal use of the Claude Agent SDK; no separate API credits are needed. Full technical details are in AGENTS.md.

Highlighted Details

• Automatic conversation capture and knowledge extraction via Claude Code hooks and Claude Agent SDK.
• Compilation into structured, cross-referenced knowledge articles.
• RAG-less retrieval for personal knowledge bases (~500 articles) using a simple index.
• Automated daily log compilation and optional manual compilation.
• lint.py for knowledge base health checks (links, orphans, contradictions, staleness).

Maintenance & Community

No specific details on maintainers, community channels, or project roadmap were found in the provided README text.

Licensing & Compatibility

The README text does not specify a software license. Compatibility for commercial use or closed-source linking cannot be determined.

Limitations & Caveats

The RAG-less retrieval is effective for personal knowledge bases up to ~500 articles; RAG is recommended for larger datasets. Functionality depends on an active Claude subscription for Claude Agent SDK access.