Hey HN, we want to share HelixDB, a native Graph-Vector Database we are building to solve the 'index rot' that plagues long-term agent memory.

Why a hybrid? Vector databases are excellent for initial semantic search, but as datasets grow, retrieval quality inevitably drifts. Why settle for just similarity when your agents need to understand the structural relationships between facts? HelixDB allows you to combine approximate vector search, BM25 full-text search, and property graph traversal in one native system.

Here are some ways teams are using HelixDB to solve complex retrieval problems:

• Entity Resolution in Multi-Agent Systems: By mapping agents as nodes and their interactions as edges, you can retrieve the 'social' context of a task rather than just fragmented text chunks.
• Legal/Compliance Auditing: Link specific document paragraphs to regulatory nodes; this allows for querying 'What rules apply to this document?' with guaranteed precision that vector search alone lacks.
• Supply Chain Digital Twins: Vectorize product specifications while using graph edges to maintain the physical hierarchy of parts, enabling multi-hop reasoning across complex bills of materials.

Many people have a thing against “yet another database engine”... but we went ahead and built HelixDB in Rust using a custom LSM-based storage engine because we believe standard RAG pipelines are structurally broken. When you stitch together a separate graph and vector store, you introduce network latency and synchronization bugs that break production environments. Our unified architecture eliminates the glue code and ensures consistency.

In terms of performance, 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 standard implementations on Neo4j for deep-link queries. You can see our initial performance report here: benchmarks.

If you’d like to try it out in a simple RAG demo, you can follow this quick-start guide: get started. We are currently in active development, and we would love to hear your thoughts. Many thanks! Comments and feedback welcome!