Repolode
Submit
zilliztech

zilliztech/memsearch

A persistent semantic memory layer for AI coding agents, storing memories as Markdown files indexed in Milvus for cross-agent recall.

View on GitHubWebsite
Stars
2,775
Forks
272
Open issues
267
Last commit
4 hours ago

What is it?

memsearch gives AI coding agents long-term memory that persists across sessions and is shared between tools. Conversation turns are summarized into daily Markdown files, which serve as the source of truth. A Milvus index is built on top of those files so the agent can run semantic and keyword-based search over past work, such as earlier decisions or debugging sessions.

Plugins are available for Claude Code, Codex, DeepSeek Harness, OpenClaw and OpenCode. A conversation in one agent becomes searchable context in the others. Developers can also use the CLI and Python API to add the same memory engine to their own agents or RAG pipelines.

The default embedding is a local ONNX bge-m3 model that needs no API key, and Milvus Lite works out of the box. Users can switch to OpenAI, Ollama or Zilliz Cloud through configuration.

Key features

  • 01Stores memories as human-editable Markdown files, with Milvus as a rebuildable index
  • 02Hybrid search combining dense vectors, BM25 and RRF reranking
  • 03Plugins for Claude Code, Codex, DeepSeek Harness, OpenClaw and OpenCode
  • 04Three-layer recall: search, expand a section, then view the original transcript
  • 05File watcher re-indexes changed files and skips unchanged chunks via SHA-256 hashing
  • 06Python API and CLI for building memory into custom agents

Use cases

  • Recalling why an architecture or library decision was made in an earlier session
  • Resuming a debugging thread by finding how a similar issue was fixed before
  • Sharing project context between Claude Code, Codex and other agents
  • Adding persistent memory to a custom agent built with the Python API

Quick start

  1. 1

    Add the memsearch marketplace in Claude Code

    /plugin marketplace add zilliztech/memsearch
  2. 2

    Install the memsearch plugin

    /plugin install memsearch
  3. 3

    Restart Claude Code to activate the plugin

Things to know

  • The default local embedding model (about 558 MB) is downloaded from HuggingFace Hub on first launch.
  • Cloud embedding providers such as OpenAI need an API key, and Zilliz Cloud or a self-hosted Milvus server is needed for multi-user setups.

Topics

#agent-memory#semantic-search#milvus#vector-search#claude-code#markdown#rag#long-term-memory#agent#claude-code-plugin

FAQ

What is memsearch?
A persistent semantic memory layer for AI coding agents, storing memories as Markdown files indexed in Milvus for cross-agent recall. memsearch gives AI coding agents long-term memory that persists across sessions and is shared between tools. Conversation turns are summarized into daily Markdown files, which serve as the source of truth. A Milvus index is built on top of those files so the agent can run semantic and keyword-based search over past work, such as earlier decisions or debugging sessions.
Is memsearch free to use? What license does it have?
memsearch is open source under the MIT license. Check the license terms before using it in commercial products.
Is memsearch still maintained?
Actively maintained. The last commit to memsearch was on Oct 22, 2026.
What can I use memsearch for?
Recalling why an architecture or library decision was made in an earlier session Resuming a debugging thread by finding how a similar issue was fixed before Sharing project context between Claude Code, Codex and other agents Adding persistent memory to a custom agent built with the Python API
How do I install memsearch?
Add the memsearch marketplace in Claude Code: /plugin marketplace add zilliztech/memsearch Install the memsearch plugin: /plugin install memsearch Restart Claude Code to activate the plugin
Where can I get memsearch?
The source code and installation instructions are on GitHub: https://github.com/zilliztech/memsearch

Summary generated by AI from the project's README. GitHub stats refresh daily.

Related projects