Megatron

Semantic memory system for Claude Code. Indexes conversation transcripts into project-local vector databases for context persistence across sessions.

Installation

claude mcp add megatron-mcp uvx megatron-mcp

That's it. Restart Claude Code and the memory tools are available.

Alternative: Manual Setup

If you prefer manual configuration, add to your .mcp.json:

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

Available Tools

Once configured, Claude Code gains these MCP tools:

Skills (Slash Commands)

If you copied the skills, these commands become available:

• /recall <topic> - Search memory for past context
• /remember - Save something to memory
• /resume - Continue where you left off
• /stats - Show memory statistics
• /sync - Sync current session to memory

CLI Commands

uv run megatron init                    # Initialize project memory
uv run megatron sync                    # Sync from latest transcript
uv run megatron project-search "query"  # Search project memory
uv run megatron resume                  # Get resume context
uv run megatron project-stats           # Show memory stats

Project Structure

When used, each project gets a .megatron/ directory containing:

• index/ - ChromaDB vector database
• state.json - Session state snapshots
• queue/ - Pending sync items

How It Works

1. Parses Claude Code JSONL transcripts from ~/.claude/projects/
2. Groups messages into user-assistant exchanges
3. Indexes chunks into ChromaDB with semantic embeddings
4. Exposes search/remember/sync tools via MCP protocol

See CLAUDE.md for detailed architecture.