ruvector

RuVector provides a persistent, adaptive memory substrate for AI agents that need to remember across sessions. The system combines local semantic embeddings (using all-MiniLM-L6-v2 by default), persistent vector storage with HNSW and flat indexes, graph and hypergraph relationships, explicit feedback learning via SONA MicroLoRA and EWC++ consolidation, and memory lifecycle controls including compaction, snapshots, and replication. Memory is organized into classes: working/session, episodic/Reflexion, semantic, procedural, causal/relational, learning, shared, and auditable. The core VectorDB offers durable vector storage with metadata filtering, collections/aliases for namespace isolation, and multiple retrieval modes including dense similarity, sparse-dense hybrid, predicate-aware ANN, temporal decay, coherence gating, graph reconstruction, multi-vector MaxSim, GNN reranking, Matryoshka funnel, and disk-backed ANN. Deployment surfaces include npm package (ruvector) for Node.js/TypeScript, Rust crate (ruvector-core), WASM package (@ruvector/wasm) for browsers, HTTP service (ruvector-server), PostgreSQL extension (ruvector-postgres), RVF cognitive containers, and optional hosted Shared Brain (mcp-brain). Native binaries cover Linux x64/arm64, macOS x64/arm64, and Windows x64. Key capabilities: local embeddings without per-query API fees, batch/parallel embedding, HNSW/flat indexes, LSM-based high-write ingestion (ruvector-lsm-ann), graph storage (ruvector-graph), compression via product quantization and RaBitQ, capability-gated retrieval, tamper-evident lineage via RVF witnesses, replication primitives, and Raft components. The monorepo includes specialized crates for temporal coherence, hybrid search, ACORN predicate-aware ANN, MaxSim, GNN reranking, Matryoshka, DiskANN, agent memory compaction, temporal tensor codecs, graph condensation, snapshots, copy-on-write branches, cache consistency, and capability-gated ANN. Known boundaries: unified AgenticMemory lacks native save/load and cross-type consolidation; metadata filtering narrows candidate sets; HNSW cold start rebuilds index; temporal coherence uses exact pairwise graph; compaction not wired into default paths; incremental snapshots and cloud backends incomplete; replication/Raft not production-ready; GNN/MRAgent/Darwin are research surfaces; RVF and PostgreSQL require separate toolchains. Security: embeddings treated as sensitive derivatives; collections not authorization boundaries; RVF witnesses tamper-evident but not encrypted; deletes don't cascade to snapshots/replicas; Shared Brain is hosted plane requiring review; pin models for offline deployments; separate tool execution from retrieval. Benchmarks included for core search, LSM writes, temporal coherence, capability-gated retrieval, and Matryoshka retrieval. Requires Rust 1.77+. MIT licensed.