RuVector is a Rust library for agent memory and fast local decisions. It combines local semantic embeddings, persistent vector retrieval, graph relationships, feedback-based learning, memory lifecycle controls, and optional shared memory. Retrieval runs locally by default with no database server or API key, and learning happens only from recorded outcomes and feedback, never from reads alone. Hosted services are optional and form a separate data boundary.
Memory follows a loop: capture an event or fact, embed it with a local ONNX model or an external provider, persist the vector with metadata and relationships, recall it by similarity, filters, time, or graph traversal, then record the outcome so ranking or learned state can adapt. Two typed layers sit over the general-purpose store. AgenticMemory combines working, episodic, semantic, and procedural memory behind one runtime API, and AgenticDB persists Reflexion episodes, skills, causal edges, policy state, session turns, and a hash-linked witness log. The application still decides what to remember, which evidence to trust, and when memories expire.
A separate decision package, @ruvector/typesafe, turns text into typed choice and score decisions using local embeddings, with confidence and abstention, for bounded tasks such as ticket routing and intent classification where a full language model call is unnecessary. On a 150-ticket test split it reports 4–10 ms p95 and 77.3–84.0% department accuracy.
The project is a monorepo, and installing the main package does not activate every crate. Several capabilities are marked as research or incomplete: the unified AgenticMemory manager has no native save and load yet, replication and Raft are primitives rather than a full network plane, and GNN reranking and graph reconstruction are not applied automatically in search. RuVector ships for Node.js and TypeScript, Rust, the browser through WASM, an HTTP service, a PostgreSQL extension, and Python bindings with a CLI, an MCP server, and LangChain, LlamaIndex, and Agentforce integrations.
Features
- Local semantic memory: remember and recall facts by meaning from the CLI or code, with a local ONNX embedding model
- Durable vector store: vectors, metadata, deletes, and restart recovery, with HNSW and flat indexes
- Typed agent memory: working, episodic, semantic, and procedural memory, plus persisted Reflexion episodes, skills, and causal edges
- Graph memory: graph and hypergraph storage with Cypher paths for multi-hop recall
- Retrieval options: sparse and dense fusion, predicate-aware ANN, temporal decay, multi-vector MaxSim, and disk-backed ANN
- Learning from feedback: SONA MicroLoRA adapters, EWC++ consolidation, and outcome-aware routing driven by recorded rewards
- Compression and recovery: product quantization, RaBitQ, compaction, full snapshots, and copy-on-write branches
- Governance: namespace isolation, capability-gated retrieval, and tamper-evident hash-linked lineage
- Typed decisions: a separate package that returns choice and score decisions with confidence from local embeddings
- MCP server: exposes memory tools to agents, with allow and deny tool policies
- Deployment surfaces: Node.js, Rust, browser WASM, an HTTP server, a PostgreSQL extension, and Python bindings
- Shared memory: an optional hosted service for cross-agent contributions, provenance, and voting
