The graph visualisation is the unusual part: most memory servers are a black box you can only interrogate one question at a time, and seeing the clusters tells you whether the thing is learning anything useful. Versioning on update is the other good decision - a corrected memory does not erase what it replaced.
A lightweight retrieval-augmented memory that runs on your machine. Notes, ideas and session context are embedded and kept between runs, with a visual graph view of how entries cluster.
- Recording user data, ideas and context as memories
- Retrieval by keyword or topic over everything stored, and a quick fetch of the most recent entries
- Updating an entry by appending to it or modifying it, with older versions kept so there is a history
- Importing a document straight into memory, and deleting entries you no longer want
- An interactive graph visualisation showing how memories cluster and relate
- Guidance delivered as resources, teaching the model when to recall past interactions and when not to
The uv runtime, and a Google API key in GOOGLE_API_KEY - used only for the Gemini embedding endpoint, which the author notes is free to use. Clients launch it with uvx and the package name; VS Code has a one-click install link, and Cursor and Cline take the same JSON block. The first run pulls dependencies and takes about a minute before it becomes responsive.
One command plus a key — uvx -q memory-plus@latest, then supply credentials