A Markdown-first memory system for AI agents powered by Milvus, featuring semantic search, multiple embedding providers, and framework integrations.
memsearch is an AI agent memory system by Zilliz that uses Markdown files as the source of truth and Milvus vector database as derived indexes. Core Design Philosophy: Markdown as human-readable, Git-friendly, zero vendor lock-in data source; vector database as rebuildable derived indexes.
Core Features:
- Semantic Search: Milvus-powered vector similarity search with hybrid search (dense vector + BM25) and RRF reranking
- Intelligent Deduplication: SHA-256 content hash prevents re-embedding unchanged content
- Real-time Sync: File watcher auto-indexes changes (create/edit/delete) with 1500ms debounce
- LLM Compression: Compresses indexed chunks into concise Markdown summaries
Embedding Provider Support: OpenAI (default), Google Gemini, Voyage AI, Ollama (local), sentence-transformers (local)
Framework Integrations: LangChain (BaseRetriever), LangGraph (ReAct agent tool), LlamaIndex (custom retriever), CrewAI (agent tool)
Installation:
pip install memsearch
# Optional providers
pip install memsearch[google] # Google Gemini
pip install memsearch[ollama] # Ollama local
pip install memsearch[local] # sentence-transformers
CLI Quick Start:
memsearch config init # Initialize config
memsearch index ./memory/ # Index Markdown files
memsearch search "query text" # Semantic search
memsearch watch ./memory/ # Real-time monitoring
Python API:
from memsearch import MemSearch
mem = MemSearch(paths=["./memory"])
await mem.index()
results = await mem.search("query", top_k=3)
Milvus Deployment Modes: Lite (local dev), Server (self-hosted production), Cloud (managed)
Claude Code Plugin: Out-of-the-box agent memory implementation with auto session summarization and semantic search.
Requirements: Python >= 3.10, cross-platform (macOS, Linux, Windows via WSL2)