engram-mcp

MCP server for Engram — persistent memory for AI.

Give Claude memory that persists across conversations.

Install

npm install -g engram-mcp

Setup

1. Get your API key

Sign up at ngrm.ai and copy your API key.

2. Configure MCP

Cursor — Edit ~/.cursor/mcp.json:

{
  "mcpServers": {
    "engram": {
      "command": "engram-mcp",
      "env": {
        "ENGRAM_API_KEY": "eng_your_key_here"
      }
    }
  }
}

Claude Code — Edit ~/.claude/mcp.json:

{
  "mcpServers": {
    "engram": {
      "command": "engram-mcp",
      "env": {
        "ENGRAM_API_KEY": "eng_your_key_here"
      }
    }
  }
}

3. Restart your IDE

Restart Cursor or Claude Code. You'll now have memory tools available.

Tools

| Tool | Description | |------|-------------| | memory_observe | Call after every user message. Learns automatically. | | memory_recall | Search memories by query. | | memory_learn | Store something explicitly. | | memory_bg | Working memory (short-term, decays). | | memory_stats | Get memory counts. | | memory_session | Start session — gets context in one call. | | memory_end_session | End session, save state. | | memory_deep | Deep exploration of a topic. |

Usage

Add rules to make Claude use memory automatically. Example:

On every message:
1. Call memory_observe with the user's message
2. Call memory_recall for relevant context
3. Respond
4. Call memory_learn for important new information

Or just ask Claude to remember things — it'll figure it out.

Frames

Memories are organized into frames:

• SELF — Identity, who the AI is
• USER — Facts about the user
• EVENTS — What happened
• KNOWLEDGE — Facts, information
• PREFERENCES — Likes, values
• GOALS — Objectives
• QUESTIONS — Open unknowns
• SKILLS — How to do things
• CONCEPTS — Patterns, theories
• RELATIONS — Connections between things

Use frames to filter recall: memory_recall("preferences", frames: ["PREFERENCES"])

Links

• Engram Dashboard
• API Documentation
• Support