Local MCP Server

Unforget provides a local MCP server that runs entirely on your machine with an embedded PostgreSQL database. No external server, API keys, or database setup required.

This is ideal for:

Claude Code / Claude Desktop — give Claude long-term memory across conversations
Cursor — remember project context, decisions, and preferences
VS Code Copilot — persistent memory across coding sessions
Any MCP client — standard MCP protocol over stdio or SSE

How It Works

Running unforget-mcp starts the full Unforget memory system locally:

Starts an embedded PostgreSQL (pgserver) automatically
Enables pgvector for semantic search
Loads local embedding model (all-MiniLM-L6-v2) — no API calls needed
Exposes 5 memory tools via MCP protocol
Data persists in ~/.unforget/mcp-data/ across restarts

No LLM calls on write. No API keys. No Docker. Everything runs locally.

Setup

1. Install the server


pipx install unforget-mcp

2. Connect your MCP client

Claude Code


claude mcp add unforget -- unforget-mcp

Or add to your project’s .mcp.json:


{
  "mcpServers": {
    "unforget": {
      "command": "unforget-mcp",
      "args": []
    }
  }
}

Cursor

Add to Cursor settings → MCP Servers:


{
  "mcpServers": {
    "unforget": {
      "command": "unforget-mcp",
      "args": []
    }
  }
}

VS Code (Copilot)

Add to .vscode/mcp.json:


{
  "mcpServers": {
    "unforget": {
      "command": "unforget-mcp",
      "args": []
    }
  }
}

Claude Desktop

Add to claude_desktop_config.json:


{
  "mcpServers": {
    "unforget": {
      "command": "unforget-mcp",
      "args": []
    }
  }
}

Available Tools

Tool	Description
`memory_store`	Store a fact, preference, or decision to long-term memory
`memory_search`	Search memories using 4-channel hybrid retrieval (semantic + BM25 + entity + temporal)
`memory_forget`	Delete memories by search query or specific ID
`memory_list`	Browse all stored memories, optionally filtered by type
`memory_update`	Update a memory with new content, preserving the version history

Tool Details

memory_store


content: string    — The fact or information to remember
memory_type: string — "insight" (facts/preferences) or "event" (interactions)
tags: string[]     — Optional tags for categorization
importance: float  — 0.0 to 1.0 (default: 0.7)

memory_search


query: string — Natural language search query
limit: int    — Max results (default: 5)

Uses 4-channel retrieval: semantic similarity, BM25 keyword matching, named entity overlap, and temporal recency. Results are fused with Reciprocal Rank Fusion and reranked with a cross-encoder.

memory_forget


query: string     — Search query to find and delete matching memories
memory_id: string — Or delete a specific memory by UUID

memory_list


memory_type: string — Filter: "insight", "event", or "raw"
limit: int          — Max results (default: 20)

memory_update


memory_id: string   — UUID of the memory to update
new_content: string — Updated content (old version preserved in history)

Environment Variables

Variable	Default	Description
`UNFORGET_DATABASE_URL`	(embedded)	Override with external PostgreSQL URL

Options


# Embedded mode (default)
unforget-mcp
 
# External database
unforget-mcp --database-url postgresql://user:pass@localhost/db
 
# Custom scoping
unforget-mcp --org-id myproject --agent-id claude
 
# SSE transport (for HTTP-based MCP clients)
unforget-mcp --transport sse --port 8767

Troubleshooting

Slow first startup

The first startup downloads the embedding model (~90MB) and initializes the database. Subsequent starts are much faster (under 10 seconds).

Port conflicts

If you’re running unforget-embed (for OpenClaw) on the default port, the MCP server uses a separate data directory (~/.unforget/mcp-data/) so they don’t conflict.