🧠 What if your AI assistant could actually remember?
Graph-vector memory service that gives AI assistants durable, relational memory. AutoMem achieves human-like associative memory performance through a hybrid architecture validated by peer-reviewed research from HippoRAG 2, A-MEM, and MELODI.
The Memory Problem – Solved
Traditional vector RAG finds similar text but misses relationships. AutoMem builds knowledge graphs with 11 relationship types, enabling AI to understand not just what you decided, but why, when, and how it connects to other decisions.
Without AutoMem: “Chose PostgreSQL for reliability” – finds the memory but not the context
With AutoMem: Understands PostgreSQL PREFERS_OVER MongoDB, RELATES_TO team expertise, DERIVED_FROM boring technology principle
🎯 Key Features
- 🧠 Graph-Vector Hybrid: FalkorDB for relationships + Qdrant for semantic search
- 🔗 11 Relationship Types: From RELATES_TO to EVOLVED_INTO – true knowledge graphs
- 🎯 Automatic Enrichment: Entity extraction, auto-tagging, pattern detection
- 🔄 Intelligent Consolidation: Memories improve over time through clustering and decay
- 🔍 Hybrid Search: Vector similarity + keywords + tags + time + importance scoring
- ⚡ Sub-Second Recall: Even with 100k+ memories
- 🎙️ Voice AI Ready: SSE sidecar enables 60ms response for conversational AI
- 📊 Research-Validated: Implements HippoRAG 2, A-MEM, MELODI principles
⚡ SSE Sidecar for Voice AI
NEW: AutoMem now works with voice AI platforms through our SSE sidecar.
Traditional MCPs use stdio and monitor text output. Voice AI needs Server-Sent Events. Our SSE sidecar bridges this gap with 60ms response times – fast enough that speakers don’t notice memory retrieval.
Now compatible with:
- ElevenLabs Conversational AI agents
- ChatGPT Developer Mode (voice)
- Claude web and mobile
- Any SSE-based AI platform
🚀 Deploy in 60 Seconds
railway up
Or run locally:
git clone https://github.com/verygoodplugins/automem.git cd automem make dev
📊 Research Foundation
- HippoRAG 2 (Ohio State, 2025): Graph-vector achieves 7% better associative memory
- A-MEM (2025): Dynamic organization with Zettelkasten principles
- MELODI (DeepMind, 2024): 8x memory compression without quality loss
- ReadAgent (DeepMind, 2024): 20x context extension via episodic memory
🛠️ Technical Architecture
- FalkorDB: Graph database for relationships and consolidation
- Qdrant: Vector database for semantic search (768-d embeddings)
- Flask API: REST endpoints with authentication
- Background Pipeline: Enrichment, clustering, decay, consolidation
- SSE Transport: HTTP POST requests with Server-Sent Events responses
🔌 Integration Options
MCP Bridge: Works with Claude Desktop, Cursor, any MCP client
Direct API: Any language, any framework
SSE Sidecar: Voice AI platforms and streaming applications
🚀 Recent Updates
- SSE Sidecar Release: Voice AI platform support with 60ms response times
- Pattern Detection: Automatic discovery of emerging themes across memories
- Creative Associations: Hourly process discovers surprising connections
- Cluster Consolidation: Groups similar memories, generates meta-insights
- Production Hardening: Retry logic, health monitoring, graceful degradation
Perfect for AI developers, researchers, and teams who need their AI assistants to maintain context, learn patterns, and build knowledge graphs over time. Whether you’re building conversational AI, research assistants, or enterprise knowledge systems, AutoMem provides the memory layer your AI needs.
]]>