AutoMem is a graph-vector memory service designed to provide AI assistants with durable, relational long-term memory, addressing the common issue of AI forgetting information between sessions. It targets developers building AI assistants who require sophisticated memory capabilities beyond traditional RAG or vector databases. AutoMem offers state-of-the-art performance in conversational memory, enabling AI to learn, connect, and evolve its understanding over time.

How It Works

AutoMem employs a graph-vector hybrid architecture, utilizing FalkorDB for graph storage and Qdrant for vector embeddings. Memories are stored with rich metadata, importance scores, temporal context, and semantic embeddings. Recall is facilitated by a novel 9-component hybrid search that combines semantic similarity, keyword matching, graph traversal, temporal alignment, and tag matching. The system connects memories using 11 distinct typed relationships and learns through automatic entity extraction, pattern detection, and neuroscience-inspired consolidation cycles. This approach allows for multi-hop reasoning and a deeper understanding of conversational context, outperforming existing memory systems.

Quick Start & Requirements

• Primary install/run command:
  • Railway: railway up (recommended, deploys 3 services in 60 seconds).
  • Docker Compose: git clone https://github.com/verygoodplugins/automem.git && cd automem && make dev (for local development).
• Non-default prerequisites: AUTOMEM_API_TOKEN (required for authentication), FALKORDB_HOST/PORT (required for graph database connection). Optional: QDRANT_URL/API_KEY for semantic search, OPENAI_API_KEY for real embeddings.
• Links: Railway Deployment Guide, MCP over SSE Sidecar Docs, API Reference, LoCoMo Benchmark Tests.

Highlighted Details

• Achieved 90.53% accuracy on the LoCoMo benchmark (ACL 2024), the academic standard for long-term conversational memory, surpassing previous state-of-the-art by 2.29 points.
• Features "Multi-Hop Bridge Discovery" to connect disparate conversation threads and understand underlying architectural philosophies or decision patterns.
• Employs a 9-component hybrid scoring system that ranks memories by meaning, combining vector, keyword, graph, temporal, tag, content, importance, confidence, and recency signals.
• Includes neuroscience-inspired memory consolidation cycles (decay, creative, cluster, forget) to maintain memory relevance and relevance over time.
• Supports 11 relationship types (e.g., RELATES_TO, LEADS_TO, PREFERS_OVER, CONTRADICTS) for building rich, structured knowledge graphs.

Maintenance & Community

AutoMem is developed by a solo developer. Community resources include the official website (automem.ai), the GitHub repository for source code and issues, and an NPM package for MCP integration.

Licensing & Compatibility

The project is released under the MIT License, allowing for free use, modification, and distribution, including for commercial purposes, with no vendor lock-in.

Limitations & Caveats

Full semantic search and the use of real embeddings require optional configuration of Qdrant and OpenAI API keys. The system necessitates proper setup and configuration of authentication tokens and database connections for operation.