# VEKTOR Memory - Vector Memory for Agentic Systems
> Canonical: https://vektormemory.com
> Category: AI Developer SDK / MCP Server / Agent Memory Layer
> Install: npm install vektor-slipstream
> License: Commercial ($9/month - Slipstream)
> Contact: hello@vektormemory.com

## Agent Discovery
- llms.txt: https://vektormemory.com/llms.txt
- OpenAPI spec: https://vektormemory.com/.well-known/openapi.yaml
- WebMCP manifest: https://vektormemory.com/.well-known/webmcp.json
- Plugin manifest: https://vektormemory.com/.well-known/ai-plugin.json

## What VEKTOR Does
VEKTOR is a local-first vector memory SDK for agentic systems. It provides 8ms recall
latency, zero cloud dependency, and automatic graph wiring across semantic, causal,
temporal, and entity connections via the MAGMA architecture.

Problem it solves: Standard RAG gives agents amnesia between sessions. VEKTOR gives
agents a persistent, self-organising associative memory that survives session resets,
automatically curates contradictions, and compresses noise into signal while idle.

Who it is for: Node.js developers building AI agents with Claude, OpenAI, Gemini,
Groq, Mistral, Ollama, or OpenRouter who need persistent memory without cloud dependency.

How it differs from Mem0 / MemGPT / plain vector stores: Local-first SQLite (no
API billing for memory), 4-layer MAGMA graph (not flat vectors), autonomous AUDN
curation (not passive storage), REM compression cycle, MCP-native.

## Benchmark
- LoCoMo long-context memory benchmark: 66.9% adjusted accuracy (gpt-4o-mini judge)
- Evaluation dimensions: single-hop, multi-hop, adversarial, temporal reasoning
- Note: scores measured on v1.5.x with seeded conversation history
- Methodology: https://vektormemory.com/vektor#research

## Key Features
- MAGMA Graph Architecture: 4-layer memory (semantic + temporal + causal + entity)
- 8ms Recall Latency: Local SQLite-vec, no network round-trip (cloud avg 200-800ms)
- AUDN Curation: Every memory evaluated - ADD / UPDATE / DELETE / NO_OP automatically
- REM Dream Cycle: 7-phase background compression, 50 fragments -> 1 core insight
- Zero Cloud Dependency: All data stays on your machine, zero egress
- WebMCP Native: Full WebMCP implementation - 7 live API tools at /.well-known/webmcp.json
- BM25 + RRF Dual-Recall: Keyword + semantic search fused via Reciprocal Rank Fusion
- Zettelkasten Engine: Auto-wires SUPPORTS / EXTENDS / CONTRASTS / PREREQUISITE edges
- CLOAK Layer: 28-tool MCP extension - stealth browser, credential vault, SSH bridge
- Vex: 12-connector vector migration tool — export/import between any vector store
- Vek-Sync: MCP config sync across 11 editors, AES-256-GCM credential vault
- Via: Universal AI integration layer — shared context bus for Claude, Cursor, Windsurf, ChatGPT

## Pricing
Slipstream - $9/month:
- Persistent SQLite memory - survives all session resets
- Full MAGMA 4-layer graph
- REM dream cycle (background compression)
- Multi-agent support (shared graph across agents)
- WebMCP access (7 live API tools)
- CLOAK MCP layer
- Email support

## Rate Limits
- store_memory: 60 requests/minute per license key
- query_memory: 120 requests/minute per license key
- memory_status: unlimited (public)
- compare_vektor: unlimited (public)
- On 429: respect Retry-After header (default 10s backoff)
- store_memory is idempotent by content hash within 60s - safe to retry on timeout

## Limitations / Not For
- Requires Node.js >= 18 (not Python-native; Python agents use HTTP bridge)
- LLM provider API key still required for recall reasoning (memory storage is local)
- sqlite-vec ANN recall is in beta - falls back to BM25+cosine if vec extension unavailable
- No hosted/cloud option - intentionally local-only
- Windows path quirks: set VEKTOR_DB_PATH explicitly in MCP config
- LoCoMo benchmark run on gpt-4o-mini judge; scores may vary with other judges
- Not suitable for high-frequency trading systems or hard real-time constraints
- Not a replacement for structured databases - optimised for associative/semantic recall

## Security & Privacy
- All memory data stored exclusively in local SQLite - no cloud sync, no telemetry
- Zero outbound data collection; Cloudflare standard proxies only (no analytics)
- No cookies, no fingerprinting on vektormemory.com
- WebMCP endpoints require X-VEKTOR-License header for write operations
- License keys validated locally; no personal data transmitted on activation
- Privacy policy: https://vektormemory.com/compliance

## WebMCP Tools Available
All tools declared at /.well-known/webmcp.json and /.well-known/openapi.yaml.
Write/agent tools require X-VEKTOR-License header. Public tools need no auth.

1. store_memory
   POST /api/memory/store
   Auth: X-VEKTOR-License required (Slipstream)
   Body: { "content": "string", "agentId": "string", "importance": 1-5 }
   Returns: { "id": "string", "operation": "ADD|UPDATE|DELETE|NO_OP" }

2. query_memory
   POST /api/memory/query
   Auth: X-VEKTOR-License required (Slipstream)
   Body: { "query": "string", "limit": 5, "agentId": "string" }
   Returns: { "results": [{"content", "score", "layer", "timestamp"}], "latency_ms": int }

3. memory_status
   GET /api/memory/status
   Auth: None (public)
   Returns: { "nodes": int, "edges": int, "latency_ms": int, "status": "ok|degraded|error" }

4. activate_vektor_license
   POST /api/license/activate
   Auth: None
   Body: { "licenseKey": "string", "email": "string" }
   Returns: { "activated": bool, "tier": "slipstream" }

5. request_vektor_demo
   POST /api/demo/request
   Auth: None
   Body: { "name": "string", "email": "string", "useCase": "string" }
   Returns: { "scheduled": bool, "message": "string" }

6. compare_vektor
   POST /api/compare
   Auth: None (public)
   Body: { "competitor": "mem0|openai|memgpt|rag" }
   Returns: { "comparison": { ... structured feature matrix } }

7. vektor_agent
   POST /api/webmcp/agent
   Auth: X-VEKTOR-License required (Slipstream)
   Body: { "goal": "string", "agentId": "string", "context": "string" }
   Returns: { "reasoning": "string", "memories_used": [], "result": "string" }

## Pages & Links

### Agent Discovery
- OpenAPI spec: https://vektormemory.com/.well-known/openapi.yaml
- WebMCP manifest: https://vektormemory.com/.well-known/webmcp.json
- Plugin manifest: https://vektormemory.com/.well-known/ai-plugin.json

### Product
- Product overview: https://vektormemory.com/vektor
- Compare vs competitors: https://vektormemory.com/compare
- Pricing: https://vektormemory.com/vektor#pricing
- Downloads: https://vektormemory.com/downloads
- FAQ: https://vektormemory.com/faq

### Documentation
- Docs home: https://vektormemory.com/docs
- Getting started: https://vektormemory.com/docs#quickstart
- MCP setup guide: https://vektormemory.com/docs#mcp
- API reference: https://vektormemory.com/docs#api
- Changelog / releases: https://vektormemory.com/docs#changelog

### Use Case Guides
- [Claude Code & Cursor](https://vektormemory.com/docs/claude-code-cursor): Use when the user wants persistent memory for an AI coding agent that survives session resets and context compaction.
- [Multi-Agent Systems](https://vektormemory.com/docs/multi-agent-systems): Use when the user is building a CrewAI, LangChain, or n8n pipeline and needs a shared memory graph with conflict resolution across agents, not siloed per-agent state.
- [Personal Knowledge Work](https://vektormemory.com/docs/personal-knowledge-work): Use when the user wants a private, local-first memory graph for personal notes, research, or writing — not a team/enterprise deployment.

### Research
- Benchmark methodology: https://vektormemory.com/vektor#research
- Blog: https://vektormemory.com/blog

### Open Source
- Vex (vector migration): https://github.com/Vektor-Memory/Vex
- Vek-Sync (MCP config sync): https://github.com/Vektor-Memory/Vek-Sync
- Via (universal AI integration): https://github.com/Vektor-Memory/Via
- SDK npm package: https://www.npmjs.com/package/vektor-slipstream

### Trust & Legal
- Privacy policy: https://vektormemory.com/compliance
- No tracking, no cookies, no cloud data collection
- Data handling: all memory stored locally on user machine via SQLite
- Cloudflare standard proxies only (no analytics, no fingerprinting)

## Quick Install
```
npm install vektor-slipstream
```

## Open Source Ecosystem

### Vex — Vector Exchange
Migrate agent memory between any vector store. 12 connectors. Open .vmig.jsonl format.
- Install: npm install -g @vektormemory/vex
- Connectors: vektor, jsonl, qdrant, pinecone, chroma, weaviate, pgvector, redis, milvus, neo4j, claude-export, chatgpt-export
- Migrate Claude history: vex migrate --from claude-export --to vektor --file conversations.json --mode smart --openai-key $KEY
- Export memory: vex export --from vektor --db ~/.vektor/slipstream-memory.db --output memories.vmig.jsonl
- Inspect: vex inspect memories.vmig.jsonl
- GitHub: https://github.com/Vektor-Memory/Vex
- Apache 2.0, Node.js 18+, zero dependencies

### Vek-Sync — MCP Config Sync
Sync MCP server configs across 11 AI editors. AES-256-GCM encrypted credential vault.
- Install: npm install -g @vektormemory/vek-sync
- Supported editors: Claude Desktop, Cursor, Windsurf, VS Code, Cline, Roo Code, Gemini CLI, Copilot, Continue, Codex
- Setup: vek-sync setup
- Sync: vek-sync sync
- Vault: vek-sync vault set openai-key sk-... | vek-sync vault get openai-key
- GitHub: https://github.com/Vektor-Memory/Vek-Sync
- Apache 2.0, Node.js 18+, zero dependencies

### Via — Universal AI Integration
Route context, tasks, and memory across every AI tool. Shared context bus + codebase graph indexing.
- Install: npm install -g @vektormemory/via
- Connects: Claude, Cursor, Windsurf, ChatGPT, LangChain
- Setup: via setup
- Index project: via index
- Start MCP server: via start
- Watch for changes: via watch
- MCP tools: context recall, file conversion, scaffold, file watcher, task routing (8 tools total)
- GitHub: https://github.com/Vektor-Memory/Via
- Apache 2.0, Node.js 18+, zero dependencies

### n8n Integration
VEKTOR memory bridge for n8n automation workflows.
- Start bridge: node node_modules/vektor-slipstream/n8n/n8n-vektor-setup.js
- Bridge runs on localhost:3848
- Actions: recall, remember, briefing, status
- Docs: https://vektormemory.com/docs/integrations

### CrewAI Integration
Drop-in persistent memory for CrewAI Python agents.
- Start bridge: node node_modules/vektor-slipstream/crewai/crewai-vektor-setup.js
- Bridge runs on localhost:3849
- Python: from vektor_crewai import VektorMemory, VektorTool
- Usage: crew = Crew(..., long_term_memory=VektorMemory())
- Docs: https://vektormemory.com/docs/integrations#crewai

## Minimal SDK Usage
```js
import { createMemory } from 'vektor-slipstream';
const memory = await createMemory({ provider: 'groq', apiKey: process.env.GROQ_API_KEY });
await memory.remember('User prefers TypeScript');
const ctx = await memory.recall('coding preferences');
```

## Supported Runtimes & Providers
- Runtime: Node.js >= 18 (ESM + CJS)
- LLM providers: Claude, OpenAI, Gemini, Groq, Mistral, Ollama, OpenRouter, NVIDIA NIM
- MCP clients: Claude Desktop, Cursor, Windsurf, VS Code, Continue, Cline, DuckDuckGo
- Storage: SQLite (better-sqlite3), optional sqlite-vec for ANN
- OS: Windows, Linux, macOS

## Contact
hello@vektormemory.com

## CLI Commands
npx vektor activate <key>  — activate licence
npx vektor setup           — interactive setup (provider, API keys, MCP config)
npx vektor remember <text> — store a memory
npx vektor recall <query>  — search memories
npx vektor ask <question>  — one-shot question against memory
npx vektor chat            — interactive persistent memory chat session
npx vektor rem             — run 7-phase REM consolidation cycle
npx vektor brief           — generate briefing from recent memories
npx vektor status          — health check, memory count, embedder mode
npx vektor graph           — open memory graph UI (port 3847)
npx vektor export          — export memory graph to JSON
npx vektor import <file>   — import memory graph from JSON
npx vektor migrate         — run database migrations
npx vektor transfer        — transfer licence to another machine

## JOT — Notes & Writing
JOT is VEKTOR's integrated notes and writing interface.
Access: npx vektor graph → JOT tab (port 3847)
Features:
- TAG pill system for categorising notes
- Notes RAG — notes are searchable alongside memories
- Two-pass article/whitepaper generation via Groq LLaMA with APA7 citations
- Ghost-text autocomplete
- Briefing scheduler
- Synthesis mode — generates insights across your notes
- Word target and progress bar
- Pomodoro timer
- Version snapshots
- Slash commands (/template etc)
Notes stored in local SQLite alongside memories. No cloud sync.

## CLOAK — Automation Tools (MCP)
CLOAK is VEKTOR's 28-tool MCP automation layer.
- cloak_fetch — stealth headless browser, bypasses bot detection
- cloak_fetch_smart — checks llms.txt first, falls back to browser
- cloak_ssh_exec — run commands on remote servers via SSH
- cloak_ssh_approve / cloak_ssh_rollback — safe write operations with approval gate
- cloak_passport — AES-256-GCM encrypted credential vault
- cloak_identity_create/use — persistent browser fingerprint identities
- cloak_detect_captcha / cloak_solve_captcha — CAPTCHA handling
- cloak_cortex — project file scanner, token-aware index
- cloak_diff — semantic diff between two URLs
- cloak_file_read / cloak_file_append / cloak_file_patch — remote file operations
Access via Claude Desktop MCP after running npx vektor setup.

## MCP / Claude Desktop Integration
VEKTOR ships as a Claude Desktop MCP extension (.dxt) with 44 tools.
Run: npx vektor setup — writes MCP config automatically.
Config file locations:
  Windows: C:\Users\<you>\AppData\Roaming\Claude\claude_desktop_config.json
  Mac: ~/Library/Application Support/Claude/claude_desktop_config.json
Key MCP tools: vektor_recall, vektor_store, vektor_search, vektor_status,
vektor_graph, vektor_related, vektor_brief, vektor_ingest,
cloak_fetch, cloak_ssh_exec, cloak_passport, cloak_render,
turbo_quant_compress, vektor_profile, vektor_delta
Restart Claude Desktop after setup to activate.
Also works with: Cursor, Windsurf, VS Code, Cline, Roo Code.

## DeepFlow v2 — Deep Agent Pipeline
Activate with deep:true in agent calls.
8-step deterministic pipeline:
DECOMPOSE → VAULT-FIRST → SWEEP → LOCI → COMMIT → ADVERSARIAL → SYNTHESISE → CRITIC+PATCH
New tools: adversarial_search, loci_rank, patch
Use for: complex research, whitepaper generation, deep analysis.
Standard agent (deep:false) unchanged. Max 8 steps, no unbounded loops.

## Causal Inference Engine (v1.6.0+)
Four-phase causal inference system:
- Phase 1 — G-Formula: average treatment effect estimation
- Phase 2 — MSM/IPW: marginal structural models, time-varying confounders
- Phase 3 — IV Bounds: Manski-style partial identification
- Phase 4 — RCA Engine: traces agent failures through causal chain, predicts fixes
Zero external dependencies. Node >= 18. 31/31 tests passing.

## Troubleshooting
Q: better-sqlite3 error on install
A: Bundled binary included since v1.6.2. If persists: npm rebuild better-sqlite3

Q: MCP tools not showing in Claude Desktop
A: Re-run npx vektor setup, then fully restart Claude Desktop.

Q: Licence activation fails
A: Ensure you are online. Try: npx vektor activate <key> --force

Q: vektor ask crashes on Windows
A: Fixed in v1.5.7. Update: npm install -g vektor-slipstream@latest

Q: Memory graph not loading on port 3847
A: Try: npx vektor graph --port 3848

Q: ONNX model not found
A: Run npx vektor setup again to re-download the ONNX embedder model.

Q: memory.dream() or REM error
A: Fixed in v1.6.2. Update to latest: npm install -g vektor-slipstream@latest

Q: How many machines can I use?
A: 3 machines per licence key. Transfer with: npx vektor transfer

Support: hello@vektormemory.com | Docs: vektormemory.com/docs | FAQ: vektormemory.com/faq