Memory Tracker MCP

An MCP server that gives an AI assistant persistent memory, backed by an OpenAI vector store.

Memories are plain text. save_memory uploads each one as a file into a vector store named MEMORIES; search_memory runs a semantic search over that store and returns the matching chunks. The store is created on first use and reused after that, so memories persist across sessions and across clients.

Requirements

• Python 3.14+
• uv
• An OpenAI API key

Setup

uv sync

Create a .env file in the project root:

OPENAI_API_KEY=sk-...

.env is gitignored. The server calls load_dotenv() at import, which resolves relative to the working directory — this is why the client configs below pass --directory.

Tools

Tool	Argument	Returns
save_memory	memory: str — the text to remember	{"status": "saved", "vector store id": ...}
search_memory	query: str — what to look for	{"results": [chunk, ...]}

Running it

Development, with the MCP Inspector:

uv run mcp dev server.py

Directly over stdio (what MCP clients do):

uv run python server.py

Client configuration

Claude Code

.mcp.json in this repo is picked up automatically when you start Claude Code in this directory. No further setup.

Claude Desktop

Add the block below to claude_desktop_config.json, then fully quit Claude Desktop (right-click the system tray icon → Quit — closing the window is not enough) and relaunch.

{
  "mcpServers": {
    "memory-tracker": {
      "command": "C:\\Users\\shivu\\.local\\bin\\uv.exe",
      "args": [
        "run",
        "--directory",
        "f:\\Agentic AI\\Memory_tracker_mcp",
        "python",
        "server.py"
      ]
    }
  }
}

Two things differ from the Claude Code config:

• Absolute path to uv.exe. Claude Desktop launches servers with a minimal PATH that usually excludes ~\.local\bin, so a bare uv fails to spawn. Claude Code inherits your shell's PATH, so the short form works there.

• Where the config file lives. For the standard installer it is %APPDATA%\Claude\claude_desktop_config.json. For the Microsoft Store (MSIX) build, AppData is redirected and the real path is:

  %LOCALAPPDATA%\Packages\Claude_pzs8sxrjxfjjc\LocalCache\Roaming\Claude\claude_desktop_config.json
  

  Editing the non-packaged path on a Store install has no effect. Reach it from the app instead via Settings → Developer → Edit Config.

Troubleshooting

Failed to build ... Expected a Python module at src\memory_tracker_mcp\__init__.py

pyproject.toml sets package = false under [tool.uv], which tells uv to treat this as a flat script project rather than build it as a package. Without it, every uv run tries to build an installable package and fails, because the server is a single server.py at the repo root and there is no src/ layout. Note that [project.scripts] still declares a memory_tracker_mcp:main entry point that does not exist — harmless while package = false is set, but it will break the build again if that line is ever removed.

Tools appear in the client but every call errors

Almost always a missing OPENAI_API_KEY. The --directory argument is what lets load_dotenv() find .env; drop it and the server still starts, but the OpenAI client has no key. As a fallback, pass the key through the config instead:

"env": { "OPENAI_API_KEY": "sk-..." }

That hardcodes the key into the config file, so prefer .env when it works.

Server shows as failed to start

Check the client's MCP log — for Claude Desktop, logs\mcp-server-memory-tracker.log in the same config directory. A spawn/ENOENT error means the uv.exe path is wrong; confirm it with where uv.

Notes

• Every save_memory call writes a temp file with delete=False and opens it without closing the handle, so temp files accumulate in %TEMP%. Passing the text directly (file=("memory.txt", memory.encode())) would avoid the temp file entirely.
• get_or_create_vector_store scans stores by name on every call, so each tool invocation costs an extra list request.