zilliztech/memsearch

A persistent, unified memory layer for all your AI agents (e.g. Claude Code, Codex), backed by Markdown and Milvus.

What it solves

AI coding agents often lose context between sessions or across different platforms. MemSearch provides a way for agents to maintain persistent, searchable memory of conversations, decisions, and workflows, ensuring that context from one session or agent is available to others.

How it works

MemSearch uses a multi-layered approach to memory:

  • Storage: It uses Markdown files as the human-readable source of truth, with Milvus acting as a rebuildable shadow index for fast retrieval.
  • Retrieval: It employs a three-layer progressive recall system (search, expand, and transcript) using hybrid search (dense vector + BM25 sparse) and RRF reranking.
  • Maintenance: It features background tasks to keep project and user notes current and can distill repetitive workflows into reusable, installable "skills."
  • Integration: It operates via plugins for existing agents (like Claude Code and OpenClaw) or via a CLI and Python API for developers building custom agents.

Who it’s for

  • Agent Users: People using AI coding tools who want their agents to remember past discussions, decisions, and preferences without manual configuration.
  • Agent Developers: Software engineers building AI agents who need to integrate persistent, semantic memory and RAG capabilities into their applications.

Highlights

  • Cross-platform sync: Shared memory across Claude Code, OpenClaw, OpenCode, and Codex CLI.
  • Hybrid Search: Combines dense vector and BM25 sparse search with RRF reranking.
  • Procedural Memory: Distills repeated workflows into reusable agent skills.
  • Local-first: Supports local embeddings via ONNX for privacy and zero cost.