I Built Self-Evolving Claude Code Memory w/ Karpathy's LLM Knowledge Bases

A developer builds a self-evolving memory system for Claude Code inspired by Andrej Karpathy's LLM knowledge base architecture. Using Claude Code hooks (session start, pre-compact, session end), the system automatically captures conversation logs, summarizes them via the Claude Agent SDK, and promotes them into a structured wiki with concepts, connections, and an index file — all stored in an Obsidian vault. Unlike Karpathy's external data ingestion approach, this system works with internal codebase conversations, creating a compounding knowledge loop where the agent gets progressively better answers over time without manual maintenance. No vector database or RAG is needed; the agent navigates markdown files using an auto-maintained index.