MemFuse v1 — Memory for AI Agents
A powerful memory system for AI agents with hybrid graph-time storage and semantic retrieval
Problem
AI agents need robust memory systems to maintain context across conversations and sessions. Traditional approaches fall short because vector-only stores lose temporal relationships, pure graph databases struggle with semantic search, and time-series alone cannot capture entity relationships.
Approach
MemFuse combines three storage paradigms in a unified system:
Hybrid Storage Architecture:
- Neo4j for entity relationships and knowledge graphs
- TimescaleDB for temporal ordering and time-based queries
- Pinecone for semantic similarity and vector search
Smart Retrieval Strategy:
Multi-dimensional memory retrieval with semantic, temporal, and graph queries.
Automatic Memory Consolidation:
Periodic background jobs merge similar memories, prune low-value entries, and strengthen important connections.
Impact
Performance metrics (after 3 months in production):
- 94% reduction in repeated questions
- 3.2x improvement in contextual relevance scores
- 40ms p95 retrieval latency at 100K+ memories
- 600+ GitHub stars, featured in AI Weekly
Developer experience:
- Type-safe API with full TypeScript support
- One-line integration
- Built-in observability dashboard
Lessons Learned
- Hybrid is better than pure: No single database paradigm handles all memory patterns well
- Pruning matters: Without cleanup, memory systems become noise generators
- Latency is critical: Retrieval must be under 100ms to keep conversations natural
Stack and Tools
- Backend: Next.js API routes, TypeScript
- Storage: Neo4j (graph), TimescaleDB (time-series), Pinecone (vectors)
- AI: OpenAI embeddings, custom ranking models
- Infrastructure: Docker, Vercel, GitHub Actions