fleet-memory/hindsight-docs/docs/sdks/mcp.md
2025-12-04 13:16:42 +01:00

1.6 KiB

sidebar_position
4

MCP Server

Model Context Protocol server for AI assistants like Claude Desktop.

Setup

The MCP server is included in the Hindsight API. When running the API with MCP enabled (default), it exposes MCP tools via SSE at /mcp/sse.

Claude Desktop Configuration

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

{
  "mcpServers": {
    "hindsight": {
      "command": "npx",
      "args": ["-y", "mcp-remote", "http://localhost:8888/mcp/sse"]
    }
  }
}

Available Tools

hindsight_put

Store a memory for a user:

{
  "name": "hindsight_put",
  "arguments": {
    "bank_id": "user_12345",
    "content": "User prefers Python for data analysis",
    "context": "programming_preferences"
  }
}

Search memories for a user:

{
  "name": "hindsight_search",
  "arguments": {
    "bank_id": "user_12345",
    "query": "What does the user do for work?"
  }
}

Usage Example

Once configured, Claude can use Hindsight naturally:

User: "Remember that I prefer morning meetings"

Claude: Uses hindsight_put

"I've noted that you prefer morning meetings."


User: "What do you know about my preferences?"

Claude: Uses hindsight_search

"Based on our conversations, you prefer morning meetings and like Python for data analysis."

Per-User Memory

The MCP tools require a bank_id for each user:

  • Each user must have a unique bank_id (user ID, email, session ID)
  • Memories are isolated by bank_id
  • Use consistent bank_id values across interactions

See MCP API Reference for full parameter details.