Memory Mcp Lite

A local-first memory layer that sits between your AI client and your filesystem, storing technical decisions, gotchas, and commands that should outlive a single session. Organizes everything in a global/project/task tree backed by SQLite with FTS5 search. Ships nine tools split across retrieval (get_global_summary, search_memory_light, get_memory_detail) and storage (remember_decision, remember_fact, upsert summaries). The retrieval discipline pushes agents through summaries first, then light search, then full detail for 1-3 hits only, keeping token burn low. Identifies projects by normalized git remote URL so memory survives directory moves. Works via npx with Windsurf, Cursor, and Claude Desktop. Point it at a local SQLite file or a remote Turso database.