Compiling Memory MCP Server Without CUDA

Added CPU Dependency Profile

1
[project.optional-dependencies]
2
cpu = [
3
    "sentence-transformers[onnx]>=2.2.2",
4
    "torch>=2.0.0"
5
]

This creates an --extra cpu installation option that:

• Forces CPU-only PyTorch without CUDA dependencies
• Uses ONNX-optimized sentence transformers for better performance
• Eliminates GPU package bloat from CPU installations

Astral UV Integration

1
[[tool.uv.index]]
2
name = "pytorch-cpu"
3
url = "https://download.pytorch.org/whl/cpu"
4
explicit = true
5

6
[tool.uv.sources]
7
torch = [
8
    { index = "pytorch-cpu", extra = "cpu" }
9
]

This sophisticated UV configuration:

• Defines a custom PyTorch CPU index that never pulls CUDA packages
• Automatically switches PyTorch to CPU-only wheels when using --extra cpu
• Ensures deterministic CPU-only builds regardless of system configuration
• Prevents accidental CUDA package installation

Usage with UV

1
# CPU-optimized installation (recommended)
2
uv sync --extra cpu

Or,

1
uv pip install ".[cpu]"

This configuration eliminates the common issue where pip install torch automatically pulls CUDA-enabled wheels even on CPU-only systems, saving disk space and preventing compatibility issues.

1. Clone the Repository

Clone the latest MCP Memory Service code:

1
git clone https://github.com/jpmrblood/mcp-memory-service.git
2
cd mcp-memory-service

2. Create Virtual Environment

Set up an isolated Python environment:

1
uv venv
2
source venv/bin/activate

3. Install CPU-Only Dependencies

Install dependencies using Astral UV for optimized CPU-only compilation:

1
# Install with CPU profile using uv
2
uv sync --extra cpu

Or/then,

1
# For compiling, enable CPU installation
2
uv pip install ".[cpu]"

4. Verify CPU Installation

Ensure the installation uses CPU-only libraries:

1
python3 -c "import torch; print(f'CUDA available: {torch.cuda.is_available()}'); print(f'PyTorch version: {torch.__version__}')"

Expected output should show CUDA available: False.

5. Download Model for Embedding

Let’s install hugging-face cli:

1
uv pip install -U "huggingface_hub[cli]"

Download the file:

1
uv run hf download sentence-transformers/all-MiniLM-L6-v2

7. Create a running script

I want to make this running as a service because I have multiple free clients that burn tokens easily.

In ~/.local/bin/run-mcp-memory.sh

1
#!/bin/bash
2
# run-mcp-memory
3
export MCP_HTTP_ENABLED=true
4
export MCP_SERVER_HOST=0.0.0.0
5
export MCP_SERVER_PORT=8000
6
export MCP_MEMORY_STORAGE_BACKEND=sqlite_vec
7
export MCP_COMMAND="uv --directory=$HOME/mcp-memory-service run memory server"
8

9
npx -y supergateway --stdio "${MCP_COMMAND}" -outputTransport streamableHttp --port 8000

Yes, I run this using supergateway, a gateway that converts stdio MCPs into HTTP streaming and SSE.

Make the script executable:

1
chmod +x ~/.local/bin/run-mcp-memory.sh

It can run well now.

7. Run as server

1
run-mcp-memory.sh &

6. Install to the Editors

1
gemini mcp add -e sse -s user memory http://localhost:8000/sse