Hey HN, we want to share HelixDB helix db, a project we are building to solve the 'memory' problem for LLM agents. Why a hybrid? Modern AI agents struggle with long-horizon reasoning because they rely on fragmented storage; they use vector databases for similarity but lack the structural relationships provided by graphs. This leads to noise-heavy retrieval. HelixDB solves this by natively combining graph and vector types.

Our benchmarking shows that we are on par with Pinecone and Qdrant for vector operations, while our graph traversal is up to three orders of magnitude faster than Neo4j, providing a significant performance leap for complex multi-hop queries.

• Indexing codebases: you can vectorize code-snippets for semantic search while using graph relationships to map function dependencies across files.
• Molecule discovery: model complex biological interactions by combining structural graph paths with vector-based molecular property embeddings.
• Personal assistant memory: store episodic events in a graph to track temporal causality while using vectors to recall vague user preferences.

Some might ask: why build 'yet another database' instead of just connecting existing ones? We think that approach creates brittle pipelines. We went ahead and built a native Rust engine because it eliminates the synchronization overhead of maintaining separate stores, which is often the biggest bottleneck in production AI apps. Check out our technical deep dive for more architectural details.

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