We Built A Long-term Memory for Your AI Agent

Most agent memory systems have gotten good at managing context. But retrieval is still the weak link — vector similarity alone breaks when you need exact names, specific dates, or structured entity relationships. How to fix this? 🔷 Hybrid retrieval — combining dense search, sparse keyword matching (BM25), and reranking (RRF) so structural relationships aren't lost 🔷 Content-level deduplication with SHA-256 hashing, so agents stop re-embedding unchanged information 🔷 Three-layer recall (summary → section → full transcript) that compresses old memories instead of accumulating everything forever That's what we built with memsearch, an open-source memory layer for AI coding agents. It works across Claude Code, OpenClaw, and more — because memory should outlive any single agent. Try it on GitHub: https://github.com/zilliztech/memsearch