t6_mem0_v2/MCP_SETUP.md

# T6 Mem0 v2 MCP Server Setup

## ✅ MCP Server Test Results

The MCP server has been tested and is working correctly:
- ✓ Server initialized successfully
- ✓ Memory instance connected (Supabase + Neo4j)
- ✓ 7 MCP tools available and functional

## Configuration for Claude Code

Add this to your Claude Code MCP configuration file:

### Location
- **macOS/Linux**: `~/.config/claude-code/mcp.json`
- **Windows**: `%APPDATA%\claude-code\mcp.json`

### Configuration

```json
{
  "mcpServers": {
    "t6-mem0": {
      "command": "/home/klas/mem0/start-mcp-server.sh",
      "description": "T6 Mem0 v2 - Memory management with Supabase + Neo4j",
      "env": {}
    }
  }
}
```

If you already have other MCP servers configured, just add the `"t6-mem0"` section to the existing `"mcpServers"` object.

## Available MCP Tools

Once configured, you'll have access to these 7 memory management tools:

### 1. add_memory
Add new memory from conversation messages. Extracts and stores important information.

**Required**: `messages` (array of `{role, content}` objects)
**Optional**: `user_id`, `agent_id`, `metadata`

**Example**:
```javascript
{
  "messages": [
    {"role": "user", "content": "I love pizza and pasta"},
    {"role": "assistant", "content": "Great! I'll remember that."}
  ],
  "user_id": "user123"
}
```

### 2. search_memories
Search memories by semantic similarity. Find relevant memories based on a query.

**Required**: `query` (string)
**Optional**: `user_id`, `agent_id`, `limit` (default: 10, max: 50)

**Example**:
```javascript
{
  "query": "What foods does the user like?",
  "user_id": "user123",
  "limit": 5
}
```

### 3. get_memory
Get a specific memory by its ID.

**Required**: `memory_id` (string)

**Example**:
```javascript
{
  "memory_id": "894a70ed-d756-4fd6-810d-265cd99b1f99"
}
```

### 4. get_all_memories
Get all memories for a user or agent.

**Optional**: `user_id`, `agent_id`

**Example**:
```javascript
{
  "user_id": "user123"
}
```

### 5. update_memory
Update an existing memory's content.

**Required**: `memory_id` (string), `data` (string)

**Example**:
```javascript
{
  "memory_id": "894a70ed-d756-4fd6-810d-265cd99b1f99",
  "data": "User loves pizza, pasta, and Italian food"
}
```

### 6. delete_memory
Delete a specific memory by ID.

**Required**: `memory_id` (string)

**Example**:
```javascript
{
  "memory_id": "894a70ed-d756-4fd6-810d-265cd99b1f99"
}
```

### 7. delete_all_memories
Delete all memories for a user or agent. **Use with caution!**

**Optional**: `user_id`, `agent_id`

**Example**:
```javascript
{
  "user_id": "user123"
}
```

## Backend Architecture

The MCP server uses the same backend as the REST API:

- **Vector Store**: Supabase (PostgreSQL with pgvector)
- **Graph Store**: Neo4j (for relationships)
- **LLM**: OpenAI GPT-4o-mini (for extraction)
- **Embeddings**: OpenAI text-embedding-3-small

All fixes applied to the REST API (serialization bug fixes) are also active in the MCP server.

## Testing

To test the MCP server manually:
```bash
cd /home/klas/mem0
python3 test-mcp-tools.py
```

This will verify:
- MCP server initialization
- Memory backend connection
- All 7 tools are properly registered

## Activation

After adding the configuration:
1. **Restart Claude Code** to load the new MCP server
2. The server will start automatically when Claude Code launches
3. Tools will be available in your conversations

You can verify the server is running by checking for the `t6-mem0` server in Claude Code's MCP server status.

## Troubleshooting

If the server doesn't start:
1. Check that all dependencies are installed: `pip install -r requirements.txt`
2. Verify Docker containers are running: `docker ps | grep mem0`
3. Check logs: Run the test script to see detailed error messages
4. Ensure Neo4j and Supabase are accessible from the host machine

---

**Last tested**: 2025-10-15
**Status**: ✅ All tests passing