Persistent memory layer that solves the forget-versus-context-explosion problem in AI coding sessions. Every tool call gets ingested, summarized through 14 content-aware compressors, and written into a local SQLite store with a synced markdown vault you can browse in Obsidian. Raw outputs compress 99% (365 KB to 3.2 KB tested) while maintaining 98%+ retrieval recall through hybrid BM25 plus vector search with optional LLM reranking. Exposes 44 tools for querying entities, topics, sessions, and knowledge pages. Auto-configures for Claude Code, Cursor, Windsurf, Cline, Continue, and others with one init command. Fully local, no API keys. Reference implementation of Karpathy's LLM Wiki pattern with the observation pipeline built in.
claude mcp add --transport stdio jubakitiashvili-context-mem uvx context-mem