Watch the full 23 chapter tour below, then explore each area at your own pace. The complete animation lasts 4 minutes 36 seconds; the README features a separate 32 second trailer.
Open the full animation · Back to the README trailer
Each chapter opens with a full width animated explainer panel. A section specific diagram and three highlighted steps introduce the idea in a nine second loop, with a plain language explanation and practical boundaries below.
The tour uses an entirely vector based intro and conceptual illustrations inspired by the supplied RuVector Explorer. The original attached bitmap is not embedded. The visual style follows the supplied Explorer screenshot: a black instrument canvas, branching search trees, orange path traversal, mint recall hints, and compact monospace labels. All five views are included: Canopy reveals expanding branches; Hyper moves the graph focus; Tree traces layered hops; Space shows query trails; Learn highlights Recall, Search, Score, Store and Adapt in sequence. These views are conceptual SVG animations; no Three.js or scripts are required. It is an explanation, not a live benchmark or interactive explorer.
RuVector combines local retrieval, relationships and feedback.
- Remember useful context
- Find related information
- Improve with measured outcomes
A guided tour of the system • Illustrations are conceptual
Capture → encode → store → recall → act → record feedback.
- Keep facts and past decisions
- Find context for the next task
- Choose what the agent may act on
Your application decides what to remember and which evidence to trust.
An embedding model turns text into a list of numbers.
- Similar ideas sit near each other
- Different wording can still match
- Use one model and dimension per store
Local semantic mode uses a model downloaded and cached on first use.
A new question becomes a vector in the same space.
- Question: Where can data run?
- Memory: Keep inference in Canada
- Retrieve context, then let the agent reason
Similarity helps find context. It does not establish truth.
HNSW is a layered network of nearby vectors.
- Start with a few long distance hops
- Move into a denser local neighborhood
- Return the closest candidates found
Approximate search trades some recall for less search work.
Your attached explorer makes search behavior visible.
- Canopy and Tree reveal each hop
- Space shows a projected vector map
- Hyper shows branches inside a disk
The WASM engine and animated JavaScript trace are separate implementations.
Compare approximate results with an exact flat scan.
- Recall: how many true neighbors return?
- Latency: how long does a query take?
- Search breadth: how much work is allowed?
Higher search breadth usually costs more work. Measure on your data.
VectorDB persists vectors, metadata and configuration.
- Store a fact with source and tenant
- Reopen the same storagePath later
- Recall it in a new agent session
Current HNSW reopening rebuilds the index. Measure cold start time.
Use memory types to organize what the agent knows.
- Working: current task and session
- Episodic: what happened and its outcome
- Semantic: facts • Procedural: skills
AgenticDB has persistent typed records; the unified ruvllm manager is in memory.
Graph memory links people, projects, events and decisions.
- Follow who approved a decision
- Connect an outcome to its procedure
- Reconstruct context across several hops
Graph reconstruction and learned shortcuts require separate integration.
Different retrieval components answer different questions.
- Dense search finds similar meaning
- Sparse search preserves exact terms
- Filters and graph paths narrow context
Core filters apply to candidates; selective filters can return fewer results.
Temporal recall can prefer recent, supported memories.
- Let stale observations lose weight
- Keep timestamps and supporting sources
- Tune retention for the actual domain
Temporal coherence is a proof of concept for moderate memory sets.
The demo learns query hints and tunes search breadth.
- Recall similar past queries and answers
- Score against an exact reference
- Store the route and adjust ef or nprobe
This demo does not change the index or run SONA. Exact scoring is a demo aid.
SONA can adapt small weights from trajectories and rewards.
- Record what the agent tried
- Supply a success or quality signal
- Consolidate to protect useful learning
Reading memory alone does not train weights or guarantee improvement.
The separate typesafe package supports bounded decisions.
- Route a support ticket to a department
- Estimate urgency with a score
- Abstain when confidence is insufficient
Use real semantic embeddings and labeled evaluation; the hash embedder is a test double.
Specialized components reduce storage and search costs.
- Quantization uses smaller representations
- Coarse search narrows the candidate set
- Reranking checks the strongest candidates
Compression can reduce recall. Core quantization settings alone do not compress storage.
Manage memory beyond insertion and search.
- Compact according to a chosen policy
- Use snapshots and portable RVF artifacts
- Branch memory for isolated experiments
Compaction and snapshot restore are not all wired into the default VectorDB path.
Keep retrieval separate from permission to act.
- Enforce user identity and access rules
- Track sources and tamper evident lineage
- Apply retention to exports and copies too
Filters are not full authorization. Witness logs do not encrypt the data.
Choose the runtime that fits your application.
- Node.js or Rust for embedded services
- Separate WASM package for the browser
- HTTP service, PostgreSQL or RVF as needed
Installing ruvector does not activate every component in the repository.
Optional shared memory adds a separate data boundary.
- Review what leaves the local process
- Retain provenance and trust controls
- Evaluate replication for your topology
Shared Brain is optional and hosted. Replication primitives are not a complete network plane.
MCP and coding hooks expose memory operations.
- Start with an explicit tool policy
- Pin the installed package for automation
- Inspect and verify generated hook config
Read only MCP profile: RUVECTOR_MCP_PROFILE=readonly
Choose a workload and measure quality alongside cost.
- Record recall, p95 latency and memory
- Compare against a tuned fixed baseline
- Promote changes only after evaluation
Explorer timings and demo savings are not production performance guarantees.
Start with npx. Remember one fact. Recall it in a new process.
- npx ruvector
- npx ruvector info
- npm install ruvector
No database server or API key required for the local memory quick start.
Use Node.js 20 or newer for the current npm package. npx fetches and runs the CLI; npm install ruvector installs it as a project dependency.
npx ruvector
npx ruvector info
npx ruvector hooks remember --semantic --type decision "The customer requires all inference to remain in Canada."
npx ruvector hooks recall --semantic --top-k 3 "Where may customer data be processed?"Run the recall command as a new process in the same project. The acceptance test is that it retrieves the stored Canada decision. The first semantic command downloads and caches its embedding model.
- Current capability map and boundaries
- Node.js API
- SONA
- Local typed decisions
- Benchmark guide
- User supplied RuVector Explorer: index.html, ruvector_wasm.js and ruvector_wasm_bg.wasm. The JS demo stores query/answer hints and adjusts ef or nprobe; it does not train SONA or change the underlying index.
Source review: 2026-10-06. Repository README blob: 251d6ae7b6ce0fbe668224dd8eefccfb445c9cef.