@steno-ai/mcp

Persistent long-term memory for Claude. One command to set up. Works with Claude Desktop, Claude Code, Cursor, and any MCP client.

Quick Start (2 minutes)

1. Create a free Supabase project

Go to supabase.com, create a new project. Copy your:

• Project URL (looks like https://abc123.supabase.co)
• Service Role Key (in Settings > API > service_role key — NOT the anon key)

2. Get an OpenAI key

Go to platform.openai.com/api-keys, create a key.

3. Run setup

npx steno-mcp-init

This will:

• Ask for your Supabase URL, Service Role Key, and OpenAI key
• Create all database tables automatically
• Write the Claude Desktop config for you

4. Restart Claude Desktop

Quit (Cmd+Q) and reopen. Then:

• Go to Settings > General → set "Tools already loaded"
• Start chatting — Claude now has persistent memory

That's it. Your data stays in YOUR Supabase project. Nothing is shared.

What you get

| Tool | What it does | |------|-------------| | steno_remember | Stores facts, preferences, decisions, people, events | | steno_recall | Searches memory with 6-signal fusion (vector + keyword + graph + temporal + recency + salience) | | steno_flush | Forces extraction of buffered session messages | | steno_update_status | Updates priority/roadmap item status (not_started, in_progress, done, blocked) | | steno_feedback | Rates whether a recalled memory was useful | | steno_stats | Shows memory statistics | | steno_end_session | Ends current session, generates summary, starts fresh next time |

How it works

Storing memories: Every message goes through LLM extraction → entity/edge creation → temporal grounding → contextual embedding → dedup → knowledge graph update.

Recalling memories: Every query runs through 6 parallel signals fused with configurable weights. Knowledge updates are tracked — newer facts supersede older ones.

Features:

• Knowledge graph with typed entities and relationships
• Temporal reasoning (eventDate + documentDate on every fact)
• Knowledge updates (newer facts automatically supersede older ones)
• Domain-scoped entity types (vehicle, startup, project — or define your own)
• Session buffering for cross-message context
• Source chunk preservation for full-context answers

Claude Code Setup

claude mcp add steno-memory -- npx -y @steno-ai/mcp \
  --env SUPABASE_URL=https://YOUR-PROJECT.supabase.co \
  --env SUPABASE_SERVICE_ROLE_KEY=eyJ... \
  --env OPENAI_API_KEY=sk-...

Claude Desktop — Manual Setup

Add to ~/Library/Application Support/Claude/claude_desktop_config.json:

{
  "mcpServers": {
    "steno": {
      "command": "npx",
      "args": ["-y", "@steno-ai/mcp"],
      "env": {
        "SUPABASE_URL": "https://YOUR-PROJECT.supabase.co",
        "SUPABASE_SERVICE_ROLE_KEY": "eyJ...",
        "OPENAI_API_KEY": "sk-...",
        "PERPLEXITY_API_KEY": "pplx-... (optional, for cheaper embeddings)"
      }
    }
  }
}

Then run the migrations manually — see migrations folder.

Environment Variables

| Variable | Required | Description | |----------|----------|-------------| | SUPABASE_URL | Yes | Your Supabase project URL | | SUPABASE_SERVICE_ROLE_KEY | Yes | Supabase service role key (not anon key) | | OPENAI_API_KEY | Yes | For LLM extraction and embeddings | | PERPLEXITY_API_KEY | No | Cheaper embeddings ($0.03/1M tokens vs $0.13) | | STENO_SCOPE_ID | No | Scope identifier (default: "default") | | STENO_SESSION_TIMEOUT_MS | No | Session auto-end timeout in ms (default: 1800000 = 30 min) |

For Developers

Use the engine directly in your app:

npm install @steno-ai/engine @steno-ai/supabase-adapter @steno-ai/openai-adapter

import { runExtractionPipeline, search } from '@steno-ai/engine';
import { SupabaseStorageAdapter } from '@steno-ai/supabase-adapter';
import { OpenAILLMAdapter } from '@steno-ai/openai-adapter';

See @steno-ai/engine for full API docs.

Part of Steno

The memory layer for AI agents. 13 packages — engine, adapters, SDK, MCP server, and more.