Memsearch,an agent memory with md as source of truth(inspired by OpenClaw) (zilliztech.github.io)

Memsearch, inspired by the OpenClaw project, has introduced a fresh approach to agent memory by utilizing markdown files as the "single source of truth." Unlike traditional memory systems that rely on vector databases, Memsearch treats markdown content as the primary data source while using a vector store merely as a derived index. This design enhances data accessibility as users can easily edit and manage their memory in human-readable markdown format, eliminating vendor lock-in and allowing for quick reconstruction of the index if needed. The technical innovations of Memsearch include smart deduplication via SHA-256 content hashing to prevent re-embedding unchanged content, live synchronization through a file watcher, and the use of LLMs for memory summarization. It supports multiple embedding providers and offers a seamless integration with Milvus, a popular vector database system. Users can implement it within their personal knowledge bases, deployed teams, or for AI agents, empowering them with persistent, searchable memory that can automatically integrate insights into their workflows. This enhances collaboration and efficiency, making Memsearch a significant development in the AI/ML landscape for managing agent memory.