これはモデルのdriaforall/mem-agent用のMCPサーバーで、Claude DesktopやLm Studioのようなアプリに接続して、黒曜石のようなメモリシステムとやり取りできます。

サポートプラットフォーム

• macOS(メタルバックエンド)
• Linux(GPU、vLLMバックエンド搭載)

プラットフォームノート:aarch64(ARM64) Linux

• ARM64 Linuxでは、ビルド失敗を避けるためにvLLMはデフォルトでインストールされていません(安定ホイールやソースビルドは失敗しません)。
• vLLMなしでインストールが成功します;以下が可能です:
• デフォルトのOpenRouter/OpenAIパス(ローカルvLLM不要)を使用する、または
• 互換性のある x86_64 ホスト上で vLLM を実行し、クライアントをそこに向けて(agent/model.py create_vllm_client を参照)。

ランニング・インシスタント

LiteLLMプロキシ(OpenAI互換)の使用

• ローカル(例:ポート4000)でLiteLLMプロキシを動作させている場合、.envでクライアントを設定します:

• 接続性の検証:

• 次に、以下のいずれかを用いてください:

• CLI: make chat-cli

• MCP経由STDIO: make serve-mcp

• HTTP経由のMCP: make serve-mcp-http

注:ARM64 Linuxでは、vLLMの代わりにこのセットアップが推奨されています。

1. make check-uv(UVがインストールされている場合はこのステップを飛ばしてください)。
2. make install:LmStudioをMacOSにインストールします。

mem-agent-mcp

This is an MCP server for our model driaforall/mem-agent, which can be connected to apps like Claude Desktop or Lm Studio to interact with an obsidian-like memory system.

Supported Platforms

• macOS (Metal backend)
• Linux (with GPU, vLLM backend)

Platform note: aarch64 (ARM64) Linux

• On ARM64 Linux, vLLM is not installed by default to avoid build failures (no stable wheels and source builds can fail).
• Installation will succeed without vLLM; you can:
  • Use the default OpenRouter/OpenAI path (no local vLLM needed), or
  • Run vLLM on a compatible x86_64 host and point the client at it (see agent/model.py create_vllm_client).

Running Instructions

Using a LiteLLM proxy (OpenAI-compatible)

• If you have a LiteLLM proxy running locally (e.g., on port 4000), configure the client via .env:

VLLM_HOST=localhost
VLLM_PORT=4000

• Verify connectivity:

curl http://localhost:4000/v1/models

• Then use either of these:
  • CLI: make chat-cli
  • MCP over STDIO: make serve-mcp
  • MCP over HTTP: make serve-mcp-http

Note: On ARM64 Linux, this is the recommended setup instead of vLLM.

1. make check-uv (if you have uv installed, skip this step).
2. make install: Installs LmStudio on MacOS.
3. make setup: This will open a file selector and ask you to select the directory where you want to store the memory.
4. make run-agent: If you’re on macOS, this will prompt you to select the precision of the model you want to use. 4-bit is very usable as tested, and higher precision models are more reliable but slower.
5. make generate-mcp-json: Generates the mcp.json file. That will be used in the next step.
6. Instructions per app/provider:
   • Claude Desktop:
     • Copy the generated mcp.json to the where your claude_desktop.json is located, then, quit and restart Claude Desktop. Check this guide for detailed instructions.
   • Lm Studio:
     • Copy the generated mcp.json to the mcp.json of Lm Studio. Check this guide for detailed instructions. If there are problems, change the name of the model in .mlx_model_name (found in the root of this repo) from mem-agent-mlx-4bit or mem-agent-mlx-8bit to mem-agent-mlx@4bit or mem-agent-mlx@8bit respectively.

Memory Instructions

• Each memory directory should follow the structure below:

memory/
    ├── user.md
    └── entities/
        └── [entity_name_1].md
        └── [entity_name_2].md
        └── ...

• user.md is the main file that contains information about the user and their relationships, accompanied by links to the enity file in the format of [[entities/[entity_name].md]] per relationship. The link format should be followed strictly.
• entities/ is the directory that contains the entity files.
• Each entity file follows the same structure as user.md.