MCP Memory Service

# Claude Desktop Setup Guide - Windows This guide helps you configure the MCP Memory Service to work with Claude Desktop on Windows without repeated PyTorch downloads. ## Problem and Solution **Issue**: Claude Desktop was downloading PyTorch models (230MB+) on every startup due to missing offline configuration. **Solution**: Added offline mode environment variables to your Claude Desktop config to use cached models. ## What Was Fixed ✅ **Your Claude Desktop Config Updated**: - Added offline mode environment variables (`HF_HUB_OFFLINE=1`, `TRANSFORMERS_OFFLINE=1`) - Added cache path configurations - Kept your existing SQLite-vec backend setup ✅ **Verified Components Working**: - SQLite-vec database: 434 memories accessible ✅ - sentence-transformers: Loading models without downloads ✅ - Offline mode: No network requests when properly configured ✅ ## Current Working Configuration Your active config at `%APPDATA%\Claude\claude_desktop_config.json` now has: - **Backend**: SQLite-vec (single database file) - **Database**: `memory_migrated.db` with 434 memories - **Offline Mode**: Enabled to prevent downloads - **UV Package Manager**: For better dependency management ## For Other Users See `examples/claude_desktop_config_windows.json` for an anonymized template with: - SQLite-vec backend configuration (recommended) - ChromaDB alternative configuration - Offline mode settings - Performance optimizations Replace `YOUR_USERNAME` with your actual Windows username. ## Key Changes Made ### 1. Config Template Updates - Removed `PYTHONNOUSERSITE=1`, `PIP_NO_DEPENDENCIES=1`, `PIP_NO_INSTALL=1` - These were blocking access to globally installed packages ### 2. Server Path Detection Enhanced `src/mcp_memory_service/server.py`: - Better virtual environment detection - Windows-specific path handling - Global site-packages access when not blocked ### 3. Dependency Checking Improved `src/mcp_memory_service/dependency_check.py`: - Enhanced model cache detection for Windows - Better first-run detection logic - Multiple cache location checking ### 4. Storage Backend Fixes Updated both ChromaDB and SQLite-vec storage: - Fixed hardcoded Linux paths - Added offline mode configuration - Better cache path detection ## Verification After updating your Claude Desktop config: 1. **Restart Claude Desktop** completely 2. **Check the logs** - you should see: ``` ✅ All dependencies are installed DEBUG: Found cached model in C:\Users\[username]\.cache\huggingface\hub ``` 3. **No more downloads** - The 230MB PyTorch download should not occur ## Testing You can test the server directly: ```bash python scripts/run_memory_server.py --debug ``` You should see dependency checking passes and models load from cache. ## Troubleshooting If you still see downloads: 1. Verify your username in the config paths 2. Check that models exist in `%USERPROFILE%\.cache\huggingface\hub` 3. Ensure Claude Desktop has been fully restarted ## Files Modified - `examples/claude_desktop_config_template.json` - Removed blocking env vars - `examples/claude_desktop_config_windows.json` - New Windows-specific config - `src/mcp_memory_service/server.py` - Enhanced path detection - `src/mcp_memory_service/dependency_check.py` - Better cache detection - `src/mcp_memory_service/storage/sqlite_vec.py` - Fixed hardcoded paths - `src/mcp_memory_service/storage/chroma.py` - Added offline mode support