udjin-labs/mnemostack

Overview

Mnemostack provides hybrid memory for AI agents by combining vector, BM25 keyword, temporal, and graph retrieval methods. It exposes these memory capabilities through MCP, HTTP, and a Python API for flexible integration.

Best for

Best for
Developers building AI agents that need durable, hybrid memory with flexible retrieval

Use cases

• Storing and retrieving agent conversational context across sessions
• Hybrid search that merges semantic similarity with keyword matching
• Managing entity relationships and temporal recency in agent memory

How to use

Install

pip install 'mnemostack[mcp]'

Tools exposed

• exact_token
• GEMINI_API_KEY
• OLLAMA_HOST
• MNEMOSTACK_COLLECTION
• MNEMOSTACK_QDRANT_URL
• MNEMOSTACK_BM25_PATHS
• MNEMOSTACK_AUTO_RECORD_IOR
• MNEMOSTACK_VECTOR_FLOOR
• MNEMOSTACK_RERANK_MODE
• MNEMOSTACK_TOKEN_BUDGET
• MNEMOSTACK_CONFIG

Tested with

Claude Desktop, Claude Code, Cursor, ChatGPT

Notes

Mnemostack provides hybrid memory for AI agents by combining vector, BM25 keyword, temporal, and graph retrieval methods. It exposes these memory capabilities through MCP, HTTP, and a Python API for flexible integration.

4 stars on GitHub. Last updated 2026-05-18. Licensed Apache-2.0.

Use cases

• Storing and retrieving agent conversational context across sessions
• Hybrid search that merges semantic similarity with keyword matching
• Managing entity relationships and temporal recency in agent memory

Pros

• Multiple retrieval strategies improve memory accuracy and relevance
• Exposed via MCP, HTTP, and Python for broad compatibility
• Open-source Python implementation enables easy customization

Cons

• Low star count suggests limited community and maturity
• Multiple memory backends add integration complexity
• Vendor as ‘Various’ implies no single support channel

Indexed from awesome-mcp-servers-punkpeye and enriched against its public facts.