ruvnet/RuVector

What it solves

RuVector provides a high-speed, local-first substrate for AI agent memory and decision-making. It eliminates the need for external database servers or API keys for basic semantic retrieval, allowing agents to maintain persistent memory across sessions and make fast, typed decisions without needing to call a full large language model (LLM) for every task.

How it works

RuVector uses a Rust-native core to handle vector embeddings and storage. It supports local semantic embeddings (via ONNX) and persistent vector retrieval. The system is organized around a "memory loop" where agents capture events, persist them as vectors, recall them via similarity or graph relationships, and then record outcomes to adapt future behavior. It offers multiple memory classes, including working, episodic, semantic, procedural, and causal memory, and provides tools for compression (quantization), temporal decay (preferring recent memories), and graph-based reconstruction.

Who it’s for

It is designed for developers building AI agents that require low-latency local memory, embedded persistence in Node.js or Rust applications, or complex memory architectures (like Reflexion or multi-hop knowledge graphs) without relying on hosted vector databases.

Highlights

• Local-First Execution: Runs semantic embeddings and vector search locally using ONNX, reducing latency and API costs.
• Typed Decision Engine: A specialized tool for bounded tasks (e.g., intent classification) that returns confidence and abstention info faster than a full LLM.
• Comprehensive Memory Types: Supports a wide array of memory structures including episodic, procedural, and causal/graph memory.
• Extensible Persistence: Offers multiple storage backends, including a local file-based VectorDB, a PostgreSQL extension, and browser-based WASM storage.
• Causal and Relational Memory: Includes graph and hypergraph storage for explicit relationships and multi-hop recall.
• Tamper-Evident Lineage: Uses hash-linked records and witness logs to ensure memory integrity.