RuVector: Build Agent Memory with Rust and ONNX

Overview

RuVector is a Rust-based toolkit for vector search, persistent agent memory, and local decisions. Its Node.js semantic-memory path uses ONNX embeddings with the all-MiniLM-L6-v2 model, while Rust and browser/WASM packages provide other ways to use its retrieval components. It is aimed at applications that need to store context across sessions and retrieve it locally, rather than relying on a hosted vector service.

The repository is a broad monorepo, not a single database package: its separate components cover graph memory, feedback-driven learning, typed decisions, and MCP integration. Those capabilities require explicit integration; installing the root ruvector package does not activate the full platform.

Key Features

• Persistent vector storage and similarity search, with Rust and Node.js interfaces.
• Local semantic embeddings through ONNX, with external embedding providers also supported.
• Metadata filtering, HNSW and flat indexes, and optional graph-based retrieval components.
• Memory-related primitives for recording outcomes and adapting from explicit feedback; reads alone do not trigger learning.
• Separate browser/WASM, Python, PostgreSQL, and RVF build or package surfaces.
• An npm CLI and MCP server for accessing memory tools from agent workflows.

Use Cases

• An agent or coding assistant that needs to retain project decisions across sessions can use the CLI or Node.js package with a local persistent store.
• A Rust application that needs embedded vector search can build directly on ruvector-core and supply its own embeddings.
• A team prototyping browser-based vector search can evaluate the separate WASM packages without deploying a database server.
• Developers exploring graph-based or feedback-driven memory can combine the relevant crates and examples, while evaluating those research surfaces separately from the default search path.

Getting Started

For the Node.js CLI, install the package and inspect the available commands:

npm install ruvector
npx ruvector info

Semantic commands download the embedding model on first use. See the README for Rust, Python, MCP, and other setup paths.

Alternatives

• memvid: Memvid packages content, indexes, and metadata in a portable file, while RuVector offers vector search and memory components across Rust, Node.js, and WASM.
• cortex-mem: Cortex Memory focuses on durable file-based agent memories with CLI, REST, and MCP access; RuVector centers on vector search with ONNX embeddings and additional components.
• OpenViking: OpenViking organizes agent context in a browsable virtual filesystem with self-hosted or hosted options, rather than RuVector's Rust-based vector-search toolkit.

Considerations

The repository is active, with frequent commits and releases and prompt issue and pull-request handling. That activity is encouraging, but does not make every component equally mature; the README identifies research surfaces and incomplete features, including incomplete consolidation and persistence in the unified runtime memory manager, index rebuilds when opening persisted HNSW databases, and limitations in replication and snapshot restoration. The core workspace requires Rust 1.77 or newer, and the npm path requires Node.js 18 or newer. The default semantic path downloads its model on first use, so offline or regulated deployments need to prepare and pin it in advance. Review the repository's security guidance as well: metadata filters are not a complete authorization boundary, and embeddings should be treated as sensitive data.