Vector + Graph + Column Store

The Agentic Memory
for Autonomous AI

The unified database combining vectors, knowledge graphs, and structured data. The cognitive backbone your AI agents need. Microsecond retrieval, VelesQL, zero cloud dependency.

Get Started GitHub

47µs

Search Latency

In-memory, 10K/768D, k=10, Cosine

21.8ns

SIMD Distance

768D dot product • AVX2

6MB

Binary Size

Self-contained, zero dependencies

100%

Recall@10

Perfect mode 10K/128D

VelesQL

-- Vector + Graph + SQL in one query

SELECT

memory.*, similarity()

FROM

agent_memory AS memory

WHERE

vector NEAR $embedding

AND MATCH (ctx)-[:RELATES_TO]->(fact)

AND session_id = $current_session

ORDER BY similarity() DESC

LIMIT 10;

Why Agents Need More Than Vector Search

AI agents need three types of memory: semantic (what feels similar), episodic (what is factually connected), and structured (explicit knowledge). Traditional databases can't provide all three unified.

Traditional Vector DBs

Vectors only, no relationships
Can't model factual connections
50-100ms latency per query
Network round-trips add up fast
No offline, no data sovereignty
Your data on someone else's servers

VelesDB Agentic Memory

Vector + Graph + Columns unified
Complete memory for AI agents
47µs search latency (10K, in-memory)
1000x faster than cloud alternatives
Runs anywhere, works offline
Server, Browser, Mobile — your data stays local

Metric	VelesDB	Cloud Vector DBs
Search Latency (10K)	47µs *	50-100ms
10 Retrievals	~0.5ms total	500ms-1s total
Time to First Token	Instant	Noticeable delay

* In-memory HNSW index, no WAL/persistence overhead. 10K synthetic vectors, 768D, Cosine, M=32, ef_c=400. Criterion.rs, i9-14900KF, Rust 1.94.1. April 2026. At scale (SIFT1M 1M×128D): 25.1x faster than Qdrant, 189x faster than Memgraph, 270x faster than ClickHouse.

Features

The complete memory system for autonomous AI agents

Semantic Memory (Vectors)

HNSW index with SIMD acceleration. What your agent perceives as similar. Multi-Query Fusion with RRF, Average, Maximum strategies.

Episodic Memory (Graph)

Native Knowledge Graph with nodes, edges, and the MATCH clause. What your agent knows is factually connected.

VelesQL - SQL + NEAR + MATCH

Unified query language for vectors (NEAR), graphs (MATCH), and structured data. No JSON DSL to learn.

Hybrid Search

Combine BM25 full-text, vector similarity, and graph traversal in a single query. Trigram Index 22-128x faster.

SIMD Native Performance

AVX-512/AVX2/NEON auto-detection. 24.7ns dot product for 768D (31.1 Gelem/s). 4-accumulator ILP kernels.

Run Anywhere

Server, CLI, Python, TypeScript, WASM, iOS, Android, Tauri. Same core, same performance.

Metadata-Only Collections

Lightweight collections without vectors for catalogs, configs, or text-only search. Memory efficient.

Advanced Quantization

PQ with OPQ rotation (configurable codebooks), SQ8 (4x), Binary (32x), RaBitQ. ADC search with SIMD lookup tables.

SIMD Performance (768D Vectors)

21.8ns

Dot Product

35.2G elem/s

26.0ns

Euclidean

29.5G elem/s

32.4ns

Cosine

23.7G elem/s

35.2ns

Hamming (768D)

768D binary vectors

27.3ns

Jaccard

768D set similarity

Raw SIMD kernel latency (single pair, no index overhead). AVX2 4-accumulator pipeline. i9-14900KF, Rust 1.94.1, Criterion.rs. April 3, 2026.

HNSW Recall Profiles (10K/128D)

Mode	ef_search	Recall@10	Latency P50	vs v1.0
Fast	64	92.2%	36µs	NEW
Balanced	128	98.8%	57µs	NEW
Accurate	512	100%	130µs	NEW
Perfect	4096	100%	200µs	NEW

Native Rust benchmarks (no HTTP overhead). Run your own: cargo bench
In-memory HNSW index (no WAL/persistence). 10K synthetic random vectors, 128D, Cosine. M=32, ef_c=400 (auto params). Recall measured vs brute-force ground truth. i9-14900KF, 64GB DDR5, Rust 1.94.1, Criterion.rs, Windows 11 Pro. April 2026.

Use Cases

The cognitive backbone for autonomous AI

Agentic Memory

Complete memory system for autonomous agents: semantic (vectors), episodic (graph), and structured data in one unified store.

Vector + Graph + ColumnsMicrosecond retrievalSemantic + Episodic + Procedural memory

VelesQL

-- Agentic Memory: Vector + Graph unified

SELECT * FROM memories

WHERE vector NEAR $embedding

AND MATCH (a)-[:KNOWS]->(b)

LIMIT 10;

GraphRAG

Combine knowledge graph traversal with vector similarity for superior context retrieval. MATCH + NEAR in one query.

Knowledge Graph nativeMATCH clauseHybrid retrieval

VelesQL

-- GraphRAG: MATCH + NEAR in one query

SELECT doc.*, similarity()

FROM documents doc

WHERE vector NEAR $query

AND MATCH (doc)-[:CITES]->(ref)

ORDER BY similarity() DESC;

AI Desktop Apps

Build offline-capable AI applications with Tauri or Electron. Single binary, no server needed.

6MB footprintWorks offlineTauri v2 plugin

TypeScript (Tauri)

const results = await invoke('plugin:velesdb|search', {

request: { collection: 'memories', vector: embedding, topK: 10 }

});

Browser Vector Search

Run vector search directly in the browser with WASM. Privacy-first, no backend required.

WASM-nativeSIMD128 optimizedData stays local

JavaScript (WASM)

import { VectorStore } from 'velesdb-wasm';

const store = new VectorStore(768, 'cosine');

const results = store.search(query_vector, 10);

Mobile AI (iOS/Android)

Native SDKs for mobile with 32x memory compression via Binary Quantization.

UniFFI bindingsARM NEON SIMD32x compression

Swift (iOS)

let db = try VelesDatabase.open(path: "./agent_memory")

let results = try collection.search(vector: embedding, limit: 10)

Robotics & Autonomous Systems

Microsecond decision-making for real-time autonomous systems. Knowledge graph for world modeling.

<100µs latencyWorld model (Graph)Offline mandatory

Rust

// <100µs latency for real-time decisions

let context = memory.search(sensor_embedding, 5);

let world_model = graph.traverse(current_node);

On-Premises / Air-Gapped

Full data sovereignty for regulated industries. GDPR, HIPAA, PCI-DSS ready.

100% local dataNo internet requiredAudit-ready

Shell

./velesdb-server --data-dir /secure/vectors --host 127.0.0.1 --port 8080

Multi-Agent Collaboration

CRDT-based memory synchronization for collaborative AI systems. Local-first, conflict-free merge.

CRDT sync (Premium)Local-firstConflict-free

TypeScript (Premium)

// CRDT sync for collaborative agent memory

await sync.connect_peer("agent_b_address");

agent_a.memory.sync();

Comparison

The only database with Vector + Graph + Columns

Looking for Agentic Memory?

Pinecone

Vector + Graph unified, no API keys, 1000x faster locally

Qdrant

25x faster kNN search (SIFT1M), native Knowledge Graph, single binary (6MB), WASM/Mobile

Neo4j

189x faster BFS traversal, vectors + graph in one engine, microsecond latency, embedded

pgvector

Purpose-built agentic memory, 1000x faster, graph native

ChromaDB

Production Rust, Knowledge Graph, VelesQL language

Feature	🐺VelesDB	Qdrant	Milvus	Pinecone	pgvector
Architecture	Single Binary	Container	Cluster	SaaS	Postgres Ext
Search Latency (10K)	47µs *	~30ms	~20ms	~50ms	~50ms
Knowledge Graph	Native MATCH	None	None	None	None
Setup Time	< 1 min	5-10 min	30+ min	5 min	15+ min
Binary Size	6 MB	100+ MB	GBs	N/A	Extension
Query Language	SQL (VelesQL)	JSON DSL	SDK	SDK	SQL
WASM/Browser
Mobile SDK
License	VelesDB Core 1.0	Apache 2.0	Apache 2.0	Closed	PostgreSQL

* In-memory HNSW, no WAL. 10K synthetic vectors, 768D, Cosine, M=32, ef_c=400. Criterion.rs on i9-14900KF, 64GB DDR5, Rust 1.94.1. April 2026. Competitor latencies include network overhead (client-server architecture). External benchmarks (SIFT1M 1M×128D): VelesDB 25.1x faster than Qdrant (kNN@10), 189x faster than Memgraph (BFS 1-hop), 270x faster than ClickHouse (columnar filter). Full results: github.com/cyberlife-coder/velesdb-benchmarks.

Why VelesDB for Agentic Memory

Unified Vector + Graph + Columns

One database for semantic, episodic, and structured memory. No data silos.

VelesQL: SQL + NEAR + MATCH

Query vectors with NEAR, traverse graphs with MATCH, filter with SQL. All in one language.

Native GraphRAG

Knowledge graph traversal combined with vector similarity for superior retrieval.

SLM-Optimized

Designed for Small Language Models (3B-8B) running on consumer hardware.

Runs Everywhere

Server, WASM, iOS, Android, Tauri. Same agentic memory, same performance.

Multi-Agent Sync (Premium)

CRDT-based synchronization for collaborative agent memory. Local-first, conflict-free.

Get Started in 60 Seconds

Download, install, and run. No complex setup, no dependencies, no cloud accounts required.

Rust (crates.io)

cargo add velesdb-core

Quick Example

1. Start the server

velesdb-server --data-dir ./my_data

2. Create a collection

curl -X POST localhost:8080/collections \
  -H 'Content-Type: application/json' \
  -d '{"name":"agent_memory","dimension":768,"metric":"cosine"}'

3. Search with VelesQL

curl -X POST localhost:8080/query \
  -H 'Content-Type: application/json' \
  -d '{"query":"SELECT * FROM agent_memory WHERE vector NEAR $v LIMIT 10","params":{"v":[0.1,0.2]}}'

Read the Documentation View Examples on GitHub