ruvnet/RuVector

What it solves

RuVector provides a persistent, adaptive memory substrate for AI agents, allowing them to remember information across sessions without requiring a database server or external API keys. It solves the problem of agent "forgetfulness" by combining semantic embeddings, graph relationships, and explicit feedback loops to create a long-term memory system that can evolve based on outcomes.

How it works

RuVector uses a Rust-native core to manage vector storage and retrieval. It supports local semantic embeddings (via ONNX and the all-MiniLM-L6-v2 model) to enable search without per-query fees. The system organizes memory into different classes—working, episodic, semantic, procedural, and causal—and implements a "memory loop" where agents capture events, persist them as vectors, recall them via similarity or graph traversal, and then adapt their ranking or learning state based on theoutcomes of their decisions.

Who it’s for

It is designed for developers building AI agents and services in Node.js or Rust who need embedded, local-first persistent memory, as well as those building complex knowledge graphs or high-write event streams for agentic workflows.

Highlights

• Local-First Memory: Operates without a database server or API keys, using local ONNX embeddings.
• Multi-Modal Memory Types: Supports episodic, semantic, procedural, and causal memory layers.
• Advanced Retrieval: Includes HNSW and flat indexes, hybrid sparse-dense search, and temporal decay to prioritize recent memories.
• Extensible Architecture: Offers specialized crates for graph storage, PostgreSQL extensions, and memory compression (Product Quantization, RaBitQ).
• Agentic Integration: Provides an MCP (Model Context Protocol) server for easy integration with LLM agents.