Claude Memory Compiler — Evolving Knowledge Base

Claude Memory Compiler — Build Memory from Your AI Sessions

The Problem

Claude Code sessions are ephemeral — valuable decisions, debugging insights, and architectural choices are lost when the session ends. You end up re-explaining the same context over and over.

The Solution

Claude Memory Compiler hooks into your Claude Code workflow and builds a growing knowledge base automatically:

1. Session End → hooks trigger flush.py
2. Extraction → Claude Agent SDK identifies decisions, patterns, and lessons
3. Daily Logs → insights saved to timestamped daily files
4. Compilation → compile.py organizes into cross-referenced articles
5. Injection → SessionStart hook loads index.md into next session

Key Features

• Automatic capture — hooks fire on session end + pre-compaction safety net
• Knowledge extraction — AI identifies decisions, patterns, and lessons learned
• No vector database — structured index.md works better at personal scale
• Health checks — lint.py runs 7 checks (broken links, orphans, contradictions, staleness)
• Daily compilation — automatically triggers after 6 PM local time
• Query interface — search your knowledge base from the command line

Setup

# Install dependencies
uv sync

# Copy hooks config to your project
cp .claude/settings.json your-project/.claude/settings.json

# Compile knowledge base
uv run python scripts/compile.py

# Query your knowledge base
uv run python scripts/query.py "How did I handle the auth migration?"

# Run health checks
uv run python scripts/lint.py

How Knowledge Is Organized

knowledge/
  index.md          # Master index, injected into each session
  architecture/     # System design decisions
  debugging/        # Bug patterns and solutions
  conventions/      # Code style and project rules
  integrations/     # Third-party service patterns

FAQ

Q: What is Claude Memory Compiler? A: An open-source tool that automatically captures Claude Code sessions and compiles them into a structured, searchable knowledge base that persists across conversations.

Q: Does it require a vector database? A: No. It uses a structured index.md approach inspired by Karpathy's LLM Knowledge Base architecture, which works better than RAG at personal scale.

Q: How does it integrate with Claude Code? A: Via Claude Code hooks — it captures sessions automatically on end, extracts insights, and injects the knowledge index at session start.