AI BriefingZilliz Unveils 'memsearch', a Shared Memory Tool for AI Agents
Key point
It's a tool that provides a markdown-based persistent memory layer so multiple AI coding agents can share context.
Details
memsearch has been unveiled to solve the problem of AI coding agents forgetting context after a session ends or data being disconnected between tools. It was created by Zilliz, the developer of Milvus.
The core of this tool is that it treats markdown (.md) files as the Source of Truth. Memory is stored as human-readable and editable text files, and Milvus serves as a 'shadow index' derived from these files.
Key Features and Architecture:
- Layered structure: It consists of 4 stages: agent plugins, CLI/API, core search engine, and markdown source.
- Hybrid search: Search accuracy is improved by re-ranking dense vector and BM25 sparse search using RRF (Reciprocal Rank Fusion).
- 3-stage progressive search: Depending on the depth of the query, information is explored step by step in the order of 'chunk search (L1) → section expansion (L2) → full original conversation parsing (L3)'.
- Auto-capture: When a conversation ends, it is automatically summarized and recorded in a markdown file, and a real-time watcher efficiently indexes only the changed parts.
It currently supports plugins for various agents including Claude Code, OpenClaw, and Codex CLI, and is provided with a Python API for developers under the MIT license.
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.