// the find
coleam00/claude-memory-compiler
Give Claude Code a memory that evolves with your codebase. Hooks automatically capture sessions, the Claude Agent SDK extracts key decisions and lessons, and an LLM compiler organizes everything into structured, cross-referenced knowledge articles - inspired by Karpathy's LLM Knowledge Base architecture.
A Claude Code hook system that automatically captures coding sessions and compiles them into a personal knowledge base — decisions, patterns, gotchas — organized as cross-referenced markdown articles. Built on Karpathy's LLM KB architecture but aimed at your own conversation history rather than web clips. Useful for developers who want persistent memory across Claude Code sessions without managing a vector DB.
The no-RAG argument is genuinely well-reasoned: at personal scale an LLM navigating a structured index beats cosine similarity, and the system avoids the operational overhead of embeddings and a vector store entirely. The hook integration is tight — session-end and pre-compact hooks mean you don't lose context even when Claude compacts mid-task. The lint.py health checks (broken links, orphans, contradictions, staleness) are a real differentiator; most personal KB tools skip this entirely and rot silently. Using uv as the package manager is the right call for 2026 Python tooling.
The setup flow — 'tell your AI to clone the repo and set up hooks' — is a clever bootstrapping trick but fragile; if the agent misreads AGENTS.md or merges settings.json incorrectly you end up with silently broken hooks and no indication anything is wrong. The after-6PM auto-compile trigger is time-zone-naive in a way that will bite anyone running across machines or CI. There's no mention of what happens when daily logs grow large before compilation — flush.py calling the Agent SDK on a fat transcript could hit context or cost limits unexpectedly. The system only works with Claude Code; if you use other editors or the API directly, you get nothing.