Hey HN, we want to share HelixDB helix db, a project we are working on to solve the limitations of flat vector search in AI agent memory. Why a graph-vector hybrid? Vector databases are excellent for similarity, but modern agents need to navigate complex, multi-hop entity relationships that flat similarity search simply cannot handle. We built HelixDB to bridge this gap, allowing you to combine semantic similarity with relational graph traversal in a single, high-performance engine written in Rust. We designed HelixDB to be the backbone for agentic workflows. Here is where it shines:

• Complex Retrieval-Augmented Generation (RAG): Instead of just finding the most similar document, your agent can start with a vector search to identify an entity and then traverse its knowledge graph connections to pull in factual metadata, drastically reducing hallucinations.
• Fraud Detection: You can vectorize transaction logs to find anomalies while simultaneously traversing the graph to identify suspicious patterns, such as multiple accounts sharing the same physical address or device ID.
• Personalized Recommendation Engines: By combining user behavior vectors with a graph of product categories and interest hierarchies, the system provides deeper, relationship-aware recommendations that outperform pure nearest-neighbor approaches.

Many people have a thing against 'yet another query language', but we went ahead and did it anyway. We built a dynamic query model that uses a Rust or TypeScript DSL sent as inline HTTP requests. We did this because it eliminates the need for slow, traditional schema migrations, allowing for rapid iteration as your agent's reasoning logic evolves. Performance is a top priority for us. Our internal benchmarking shows that while we are on par with Pinecone and Qdrant for raw vector throughput, our native graph-traversal engine is up to three orders of magnitude faster than Neo4j for multi-hop agent queries, primarily due to our tiered caching system that keeps hot-path data ready in SSD and memory. You can see how this works in our architecture guide: introduction. We believe moving away from disjointed infrastructure—where you have to manage both a vector store and a graph database separately—is the key to building the next generation of AI agents. If you’d like to try it out in a simple RAG demo, you can follow our quick-start guide: introduction. Many thanks! Comments and feedback welcome!