Hey HN, we want to share HelixDB helix db, a project we've been building to solve the messy problem of shared memory in multi-agent systems. Why a graph-vector hybrid? Vector databases are fantastic for similarity, but as soon as you have a fleet of agents, you need to manage relationships, provenance, and concurrent updates—capabilities standard vector stores simply don't have. Our hybrid architecture bridges this gap by combining vector search with high-concurrency graph traversal.

Use Cases

• Multi-Agent RAG: Agents can collaboratively update a shared knowledge base without the risk of stale data or corruption by using our atomic graph-based locking system.
• Complex Reasoning: By combining graph edges with vector search, you can perform multi-hop reasoning across your entire knowledge graph to find connected insights that pure vector search would miss.
• Automated Provenance: Track exactly which agent created which piece of information, ensuring full auditability in production environments by mapping agent IDs to graph nodes.

Why a Custom Query Engine?

Many developers have a thing against 'yet another query language,' but we went ahead and built a custom DSL in Rust and TypeScript because it allows for dynamic HTTP-based execution, which we found significantly easier to integrate into existing agent loops than standard REST APIs. We anticipate some skepticism regarding the learning curve, but it eliminates the need for bulky SDKs and simplifies distributed deployment. You can view our source code architecture to see why we chose this specific implementation.

Performance Benchmarking

Our benchmarking shows that we are on par with Pinecone and Qdrant for pure vector throughput, but our graph-traversal engine is up to three orders of magnitude faster than Neo4j when performing high-concurrency memory lookups for agent fleets. See our latest research on memory here: 2606. For a deeper dive into our design, check out our architecture documentation.

If you'd like to try it out in a simple RAG demo, you can follow this quick-start guide. Many thanks! Comments and feedback welcome!