RuVector is a high-performance vector database written in Rust that combines approximate nearest-neighbor search with graph neural networks, so the index itself can learn and improve. Created by ruvnet, it targets real-time decisions and long-term memory for AI agents. The project is open source on GitHub with more than 4,600 stars and active development through 2026.
Standard vector stores treat embeddings as static rows to be searched. RuVector adds a learning layer: a graph neural network over the vector graph supports self-optimization of retrieval, attention mechanisms for context weighting, and features aimed at agent memory, where what an agent should recall changes over time. The Rust core focuses on low latency, and the project ships ONNX integration, a WASM build for browser and edge use, and connectors such as Neo4j synchronization. It also bundles utilities like OCR-related tooling from the same ecosystem, making it a toolkit for retrieval-heavy agent systems rather than a bare index.
RuVector exposes both a Rust API and service interfaces, and it serves as the memory layer in its author’s wider agent tooling ecosystem. Its benchmarks emphasize tail latency under concurrent load, and the graph layer supports partitioning strategies for scaling a single index across multiple shards as collections grow.
RuVector is free and open source under the MIT license. You can embed it or run it on your own infrastructure without licensing fees or usage caps.
RuVector is for developers building agent memory, semantic search and real-time recommendation systems who want a fast, hackable Rust vector engine that goes beyond a static embedding index.