Hey HN, we want to share HelixDB helix db, a project we've been building to solve the state management crisis in AI agent memory. Why a graph-vector hybrid? Vector databases are excellent for semantic similarity queries, but they often struggle when AI-driven applications need to maintain an accurate, evolving world model that requires both similarity and explicit, updatable relationships. When facts change, flat vector stores often retain contradictory data, leading to agent hallucinations. HelixDB natively integrates graph and vector types to ensure multi-hop reasoning is reliable and state updates are consistent. Here are some core use cases where this architecture shines: - Maintaining dynamic agent state: Agents can formally overwrite obsolete facts while preserving historical context through ACID-compliant transactions, preventing the 'stale data' problem. - Complex multi-hop reasoning: By traversing nodes and edges, agents can verify connections between entities, which significantly outperforms simple cosine similarity for knowledge-heavy tasks. - Enterprise knowledge graphs: You can link disparate documents via vector embeddings while simultaneously enforcing strict relational pathways for data compliance and hierarchical navigation. Many people have a thing against “yet another query language,” but we went ahead and did it anyway with our Rust/TypeScript DSL, because we think it makes working with an agent's memory graph substantially more intuitive than managing raw JSON blobs or complex SQL joins. Our benchmarking shows that we are on par with Pinecone and Qdrant for raw vector throughput, and our graph traversal capabilities are up to three orders of magnitude faster than Neo4j in high-concurrency agent environments. We’ve designed this to be fully persistent. See how we handle object storage integration here: introduction. For a deeper look at why graph-vector structures lift recall accuracy, check out this research context: agent memory retrieval vector graph 2026. If you’d like to try it out in a simple RAG demo, you can follow this guide: quickstart. Many thanks! Comments and feedback welcome!