claude-memory-mcp

Local, Persistent Memory for Any MCP-Compatible AI A lightweight, zero-cloud, token-efficient Model Context Protocol (MCP) server that gives your AI durable, searchable, and context-aware memory - entirely under your control.

Built with TypeScript, SQLite + FTS5, and minimal runtime dependencies (MCP SDK + better-sqlite3), it runs locally and stores everything in a single, portable .db file.

Works with Claude Desktop, Cursor, Windsurf, or any MCP client.

📦 Now available on npm: Install with npx @whenmoon-afk/memory-mcp

Why Local Memory?

You Control	Cloud Services
Data never leaves your machine	Sent to remote servers
Portable `.db` file	Locked in proprietary storage
Full audit & backup	Opaque retention policies
Zero recurring cost	Subscription required

Features

Feature	Benefit
Dual-Response Pattern	Returns all matches (compact index) + full details within token budget
Token Budgeting	Auto-respects `max_tokens` (~30% index, ~70% details)
Hybrid Relevance Scoring	40% relevance, 30% importance, 20% recency, 10% frequency
Auto-Summarization	Generates ≤20-word natural-language summaries
Entity Extraction	Detects people, tools, concepts, preferences
FTS5 Full-Text Search	Sub-10ms queries, Unicode, stemming — no embeddings
Provenance Tracking	Full audit trail: source, timestamp, updates
Soft Deletes	Memories preserved for debugging/rollback
Single-File DB	`memory.db` — copy, backup, move freely

Installation

Prerequisites

Node.js ≥ 18
An MCP-compatible client (Claude Desktop, Cursor, Windsurf, etc.)

Option 1: NPM Package with Auto-Setup (Recommended)

Automatic installation (configures Claude Desktop for you):

npx @whenmoon-afk/memory-mcp

This will automatically:

Detect your operating system (macOS/Windows/Linux)
Add the memory server to your Claude Desktop configuration
Create a backup of your existing config
Configure the correct command format for your platform

After installation, restart Claude Desktop completely (quit and reopen).

Or install globally:

npm install -g @whenmoon-afk/memory-mcp
memory-mcp

Option 2: From Source

For development or customization:

git clone https://github.com/WhenMoon-afk/claude-memory-mcp.git
cd claude-memory-mcp
npm install
npm run build

Output: dist/index.js — your memory server.

Integrate with Your MCP Client

Add to your client's MCP config file:

Claude Desktop:

macOS: ~/Library/Application Support/Claude/claude_desktop_config.json
Windows: %APPDATA%\Claude\claude_desktop_config.json
Linux: ~/.config/Claude/claude_desktop_config.json

Cursor/Windsurf: Check your editor's MCP settings

💡 Recommended: Use the automatic installer (npx @whenmoon-afk/memory-mcp) which handles platform differences automatically.

Manual Configuration (macOS/Linux)

{
  "mcpServers": {
    "memory": {
      "command": "npx",
      "args": ["-y", "@whenmoon-afk/memory-mcp"],
      "env": {
        "MEMORY_DB_PATH": "./memory.db"
      }
    }
  }
}

Manual Configuration (Windows)

Windows requires the cmd /c wrapper to execute npx properly:

{
  "mcpServers": {
    "memory": {
      "command": "cmd",
      "args": ["/c", "npx", "-y", "@whenmoon-afk/memory-mcp"],
      "env": {
        "MEMORY_DB_PATH": "./memory.db"
      }
    }
  }
}

Using Global Install

{
  "mcpServers": {
    "memory": {
      "command": "memory-mcp",
      "env": {
        "MEMORY_DB_PATH": "./memory.db"
      }
    }
  }
}

From Source

{
  "mcpServers": {
    "memory": {
      "command": "node",
      "args": ["/absolute/path/to/claude-memory-mcp/dist/index.js"],
      "env": {
        "MEMORY_DB_PATH": "./memory.db"
      }
    }
  }
}

Restart or reload MCP servers.

MCP Tools

Tool	Input	Description
`memory_store`	`{ content, type, importance?, entities?, ttl_days?, provenance? }`	Store or update memory with auto-summary
`memory_recall`	`{ query, type?, entities?, limit?, max_tokens? }`	Search memories with token-aware loading
`memory_forget`	`{ id, reason? }`	Soft-delete memory (preserves provenance)

Store a Preference

{
  "tool": "memory_store",
  "input": {
    "content": "User works best in focused 90-minute blocks with 15-minute breaks.",
    "type": "fact",
    "importance": 8
  }
}

→ Auto-creates:

Summary: "User follows 90/15 focused work cycles."
Entities: focused work, 90-minute blocks
Provenance: current session

Smart Recall (Dual-Response)

{
  "tool": "memory_recall",
  "input": {
    "query": "work habits",
    "max_tokens": 1200
  }
}

Response:

{
  "index": [
    { "id": "mem_8b1", "summary": "User follows 90/15 focused work cycles.", "score": 96 },
    { "id": "mem_2c9", "summary": "User avoids meetings before 10 AM.", "score": 88 }
  ],
  "details": [
    {
      "id": "mem_8b1",
      "content": "User works best in focused 90-minute blocks with 15-minute breaks.",
      "entities": ["focused work", "90-minute blocks"],
      "provenance": { "source": "chat_2025-11-03", "created_at": "2025-11-03T14:10Z" }
    }
  ],
  "meta": {
    "tokens_used": 698,
    "total_matches": 2,
    "truncated": false
  }
}

Your AI knows what it knows — and can ask for more.

Forget a Memory

{
  "tool": "memory_forget",
  "input": {
    "id": "mem_8b1",
    "reason": "Information no longer relevant"
  }
}

Response:

{
  "success": true,
  "memory_id": "mem_8b1",
  "message": "Memory soft-deleted successfully. Reason: Information no longer relevant"
}

Soft-deleted memories are preserved in the database with full provenance trail. They won't appear in searches but remain recoverable.

Dual-Response Pattern

[index]   → All matching summaries (~20 tokens each)
[details] → Full content of top memories (budget-capped)
[meta]    → tokens_used, total_matches, truncated

Index: Always included → discovery
Details: Budget-safe → precision
Follow-up: Use ids: [...] to expand

Database & Portability

File: memory.db (SQLite) — path via MEMORY_DB_PATH
Portable: Copy to USB, cloud sync, or new machine
Backup: Just copy the file
Tip: For extra security, store memory.db on a VeraCrypt Encrypted USB drive (adds friction, but maximum control).

Dependencies

This project uses minimal runtime dependencies to keep the package lightweight:

Dependency	Version	Purpose
`@modelcontextprotocol/sdk`	^1.0.4	Official MCP protocol implementation
`better-sqlite3`	^11.0.0	Fast, native SQLite3 bindings with FTS5 support

Why these dependencies?

MCP SDK: Required for implementing the Model Context Protocol standard
better-sqlite3: Native performance for full-text search and database operations, essential for memory recall speed

All other dependencies are dev-only (TypeScript, testing, linting).

Environment Variables

Var	Default	Description
`MEMORY_DB_PATH`	`./memory.db`	Database file location
`DEFAULT_TTL_DAYS`	`90`	Default time-to-live for memories (days)

Security

This is a local-only MCP server.
Data is stored in a plain SQLite file (memory.db).
For sensitive data, use OS-level encryption (FileVault, BitLocker).

Best Practices

Start with max_tokens: 1000 — adjust per model and task.
Filter by type to reduce noise and improve relevance.
Use entity filtering to narrow searches to specific topics.
Reference provenance: Track source and context for audit trails.
Backup memory.db regularly — it's just a file!

Quick Links

📦 NPM Package: https://www.npmjs.com/package/@whenmoon-afk/memory-mcp
🐙 GitHub Repository: https://github.com/WhenMoon-afk/claude-memory-mcp
🐛 Report Issues: https://github.com/WhenMoon-afk/claude-memory-mcp/issues
📖 MCP Documentation: https://modelcontextprotocol.io

Contributing

Contributions are welcome! Feel free to:

Report bugs or request features via GitHub Issues
Submit pull requests for improvements
Share your use cases and feedback

License

MIT License - see LICENSE file for details.

Built with ❤️ for the MCP community

Claude记忆