memsearch (hybrid BM25 + vector retrieval over markdown)

A lightweight hybrid search library that gives AI agents persistent memory by indexing markdown files with BM25 and vector retrieval.

Memsearch solves the context loss problem in AI agents by treating markdown files as a single source of truth for long-term memory. It uses a dual-engine approach: BM25 for precise keyword matching (essential for SKUs and specific terms) and vector embeddings for semantic recall. The system features automated SHA-256 deduplication to prevent redundant embedding costs and a live-sync watcher that re-indexes content the moment a file changes. Built by the Milvus team, it includes a ready-made Claude Code plugin to provide agents with cross-session continuity through a progressive three-level search architecture.

https://github.com/milvus-io/memsearch