MemSearch is an open-source memory layer from Zilliz that gives AI agents persistent, unified long-term memory across tools and sessions. The GitHub project has gathered around 2,770 stars and is actively developed in 2026. Instead of each coding assistant forgetting everything when a session ends, MemSearch stores memories as human-readable Markdown backed by Milvus vector search, so Claude Code, Codex, and other agents can share one memory.
The system works as a plugin for agent harnesses: conversations and working notes are chunked, embedded, and indexed in Milvus, then retrieved through hybrid semantic and keyword search when a new session needs context. Because the source of truth is plain Markdown files, your memory stays inspectable, editable, and portable rather than locked inside a proprietary database. It is designed for the emerging multi-agent workflow where a developer moves between several coding agents during one project.
MemSearch’s design reflects a simple conviction: agent memory should belong to the user, in files they can read, back up, and move between tools. Sessions are distilled into durable facts and project context rather than raw transcripts alone, which keeps retrieval relevant as memory grows. Because it builds on Milvus, the same memory store scales from a laptop experiment to a shared team deployment without changing the plugin interface your agents already use.
MemSearch is free and open source under the MIT license. Running it locally with Milvus costs nothing beyond your own infrastructure; managed Milvus cloud plans are optional.
Developers who use multiple AI coding agents and are tired of re-explaining their project to each one, and agent builders who need a portable, inspectable memory backend.