Persistent memory for AI coding agents - Open-source MCP server. #5 on LoCoMo. 100% local. MIT.

Hey Product Hunt 👋 — Vitalii here, maker of total-agent-memory. Short version: it's an open-source persistent memory layer for AI coding agents. MIT, Go, runs locally on SQLite + Postgres. Why I built it: every long Claude Code or Cursor session hits the same wall — context evaporates across sub-agents, decisions get forgotten, the agent re-asks things it knew an hour ago. Existing memory products (Mem0, Letta, Zep) are designed for chat assistants — they retrieve over conversational history. Coding agents have a different recall pattern: exact symbol/file names, fuzzy near-misses, semantic intent, all at once. So I built a 4-tier hybrid retriever: FTS5 / BM25 — lexical precision Trigram fuzzy — typo and near-miss tolerance pgvector + BAAI embeddings — dense semantic Cross-encoder rerank — final best-of-N Benchmark: 97.45% R@5 on LongMemEval (the strongest public memory benchmark). Methodology is reproducible from the repo. What's there today: MCP server, drop-in for Claude Code, self-hostable, no SaaS. Free, open source, no signup. Roadmap: managed/hosted tier for teams who don't want to run Postgres themselves, enterprise support, deeper Cursor integration. Big thanks to @anthropic / Claude Code, @cursor, and @supabase — products this is built on top of and around. Would love your feedback — especially from anyone running long-horizon coding agent sessions. What breaks for you? Where does memory fail today?