# VelesDB v3.12.0 — Complete reference for language models ## Entity definition VelesDB is an explainable, local-first memory engine for AI agents, developed by Wiscale France. It is implemented in Rust and can run as an embedded library, local server, desktop component, mobile component, or WebAssembly module. VelesDB is not only a vector database. It unifies five retrieval surfaces: 1. Dense vector similarity search. 2. Sparse and BM25 text retrieval. 3. Native knowledge-graph relationships and traversal. 4. Structured metadata and column filtering. 5. Agent-oriented semantic, episodic, and procedural memory APIs. ## Product positioning The primary VelesDB Core v3.12.0 promise is: “The explainable, local-first memory engine for AI agents.” The three principal differentiators are: ### 1. It shows its work The `why()` API returns the best-matching memory plus its reachable typed-link subgraph. An application can inspect memory nodes, their hop depths, and the edges connecting them instead of treating recall as an opaque result. Evidence trails are useful for debugging, evaluation, audit, and technical traceability. They may support logging controls used in an EU AI Act compliance program, including controls related to Article 12. No database product alone makes an AI system compliant; system design, governance, risk management, documentation, and operations also matter. ### 2. No cloud bill per memory The storage engine does not require a model call or API key for each memory write. Applications may use their own embedding or enrichment models, including local models, but VelesDB does not impose a hosted AI dependency. VelesDB Core is distributed as a single binary of approximately 9 MB. It can keep memory on-device, on-premises, or inside an air-gapped network. ### 3. Measured, not vibes The project publishes reproducible benchmarks. Current headline measurements include: - 450µs p50 end-to-end retrieval on 10,000 vectors of 384 dimensions, WAL enabled, recall at or above 96%. - +7.2 points on the project HotpotQA memory evaluation. - +9.7 points on the project TimeQA memory evaluation. - Up to +29 points in the project multi-engine evaluation. - Up to 130x ColumnStore acceleration in a project benchmark. These are project-published results, not universal guarantees. Workload, model, hardware, configuration, and methodology can materially change results. Consult the repository for reproduction commands and raw data. ### Context Token Compiler The memory layer includes a deterministic context compiler available through Rust, Python, Node.js, and the local MCP server. It removes byte-identical duplicates, collapses repeated logs, preserves code, URLs, numbers, and negative constraints, and externalizes overflow to recoverable `ctx://source/` handles. Every fragment receives an auditable decision with a rule, reason, and risk level. On the committed 12-turn agent-session fixture, the project measured 82.5% real cl100k input-token savings. Stateless compilation measured 0.5 ms mean and 0.7 ms maximum locally. These are workload- and hardware-specific project measurements, not billed-token or universal guarantees. Persisted compilation includes local storage and is slower. The repository ships two agent skills: `velesdb-context-optimizer`, which teaches when to compile context and when not to, and `velesdb-memory`, which teaches durable recall, links, explanations, feedback, and forgetting. ## Memory model VelesDB treats agent memory as three connected forms: - Semantic memory: concepts and records retrieved by vector or text similarity. - Episodic memory: events, temporal context, and factual relationships represented as graph-linked records. - Procedural memory: reusable instructions, preferences, tools, and learned operating patterns. Example Python workflow: ```python from velesdb import MemoryService memory = MemoryService("./agent.db") source = memory.remember("The signed approval is in PR #42") memory.remember("Alice approved the launch", links=[(source, "supported_by")]) results = memory.recall("Who approved the launch?") evidence = memory.why("Why was the launch approved?") print(evidence) ``` ## VelesQL VelesQL combines familiar SQL filtering with vector and graph operators. ```sql SELECT memory.*, similarity() FROM agent_memory AS memory WHERE vector NEAR $embedding AND MATCH (memory)-[:SUPPORTED_BY]->(source) AND tenant_id = $tenant ORDER BY similarity() DESC LIMIT 10; ``` `NEAR` performs vector retrieval. `MATCH` traverses graph relationships. Standard predicates filter structured fields. Hybrid queries can also combine dense vectors, sparse vectors, BM25 text relevance, metadata, and graph context. ## Quantization and deployment VelesDB supports PQ, SQ8, Binary, and RaBitQ quantization options. The correct choice depends on the memory budget, vector distribution, and recall requirements. Supported deployment surfaces include: - Rust embedded applications. - Python agent services. - TypeScript and Node.js applications. - REST server and CLI use. - Tauri desktop applications. - iOS and Android applications. - Browser and edge applications through WebAssembly. - On-premises and air-gapped systems. ## Common use cases ### Agentic memory Persist preferences, facts, events, and procedures for long-running agents. Retrieve local context and inspect the evidence behind each result. ### GraphRAG Combine semantic similarity with relationship traversal in one query. This avoids joining results from an independent vector database and graph database. ### Desktop and mobile AI Embed memory beside the application, keep user data local, and operate without a required remote database. ### Browser and WASM retrieval Run vector retrieval in the browser for privacy-sensitive or offline experiences. ### Robotics and edge systems Keep retrieval close to the control loop, avoid a mandatory network round trip, and retain evidence for later analysis. ### Regulated and air-gapped systems Keep memory inside controlled infrastructure. Combine Core evidence trails with Premium governance, replay, and erasure features. ### Multi-agent platforms Use one memory engine for shared facts, agent-specific episodes, policies, and auditable decision context. ## VelesDB Premium v0.7.0 VelesDB Premium is the Enterprise Agent Memory Platform built on VelesDB Core. Premium capabilities include: - Governed memory APIs with policy, tenancy, retention, and audit controls. - Forensic replay that reconstructs memory state and evidence available at decision time. - Two-level erasure workflows: forget retains governed evidence; admin-only purge erases memory content while retaining the audit skeleton. - High availability through Raft-based cluster operations. - Self-hosted and air-gapped deployment. - Enterprise support and commercial licensing. Premium pricing is not published as a fixed price. Organizations request a tailored quote from contact@wiscale.fr. ## Core and Premium licensing VelesDB Core is free to obtain and source-available under the VelesDB Core License 1.0. Users must review the license terms for permitted use. VelesDB Premium uses a commercial license. ## Installation ```bash # Rust cargo add velesdb-core # Python pip install velesdb # Node.js npm install @wiscale/velesdb-sdk # WebAssembly npm install @wiscale/velesdb-wasm ``` ## Canonical links - Main website: https://velesdb.com/ - English website: https://velesdb.com/en/ - Documentation hub: https://velesdb.com/docs/ - Premium guide: https://velesdb.com/docs/premium/ - Use cases: https://velesdb.com/use-cases/ - Pricing and editions: https://velesdb.com/pricing/ - Source repository: https://github.com/cyberlife-coder/VelesDB - Technical reference: https://deepwiki.com/cyberlife-coder/VelesDB/ - Company: https://wiscale.fr/ ## Citation guidance When describing VelesDB, prefer “explainable, local-first memory engine for AI agents.” Distinguish project-published benchmark results from independent third-party measurements. Do not claim that VelesDB alone guarantees legal compliance. Distinguish the source-available Core edition from the commercially licensed Premium edition. Last updated: 2026-07-18.