A persistent memory system for AI agents, memsearch allows agents to store and retrieve information semantically within human-readable Markdown files. It offers a flexible, vendor-neutral solution for developers building AI agents who need robust, easily manageable memory, integrating seamlessly into any agent framework.

How It Works

memsearch treats Markdown files as the definitive source of truth for agent memories. It indexes these files into a vector database (defaulting to Milvus Lite) using pluggable embedding providers. The system employs SHA-256 hashing for smart content deduplication, preventing re-embedding of unchanged data. A file watcher ensures live synchronization, automatically indexing new or modified files and removing stale chunks. This architecture ensures the vector store is a derived index, always rebuildable from the human-readable Markdown source.

Quick Start & Requirements

Installation is straightforward via pip: pip install memsearch. Optional embedding providers like Google Gemini, Voyage AI, Ollama, or local sentence-transformers can be installed with pip install "memsearch[provider]". The Python API involves initializing MemSearch with paths and calling index() and search(). CLI commands like memsearch index, memsearch search, and memsearch watch are available. API keys for cloud embedding providers are required, or local models must be set up (e.g., via Ollama). A ready-made Claude Code plugin is also provided.

Highlighted Details

• Markdown-first memory architecture ensures human readability and Git compatibility.
• Smart deduplication via SHA-256 content hashing optimizes storage and indexing.
• Live synchronization with a file watcher automatically updates the vector index.
• Offers a drop-in Claude Code plugin for persistent AI agent memory across sessions.
• Supports hybrid search (dense vector + BM25) with RRF reranking.
• Integrates with popular agent frameworks like LangChain, LangGraph, LlamaIndex, and CrewAI.
• Runs locally by default, prioritizing data privacy.

Maintenance & Community

Discussions and support are available via Discord. The project provides a Contributing Guide for developers interested in participation. It is inspired by OpenClaw's memory architecture and leverages Milvus for vector storage.

Licensing & Compatibility

The project is released under the MIT license, permitting commercial use and integration into closed-source applications without significant restrictions. Its local-first default enhances data privacy.

Limitations & Caveats

While robust examples are provided, specific performance benchmarks are not detailed. The effectiveness of semantic search and memory recall is dependent on the chosen embedding model and the quality/structure of the Markdown memory files.