Using it from Claude or ChatGPT¶

mimem produces a script. Two things are missing before that is useful on its own: a voice, and something that comes back tomorrow to ask you about it. An assistant cannot be a voice — that is still to come — but it can be everything else, and it can write the explanations that otherwise need a paid API key.

With this set up you can say "make me a study programme from this paper" and then, days later, "quiz me on what's due".

Set it up¶

You need Python 3.11+, uv, and the app.

uv run python packaging/build_extension.py

uvx --from git+https://github.com/jepegit/mimem --with mcp mimem-mcp

{ "command": "uv", "args": ["run", "--directory", "/path/to/mimem", "mimem-mcp"] }

uv run mimem build paper.pdf --out out/paper

Try it¶

Make a study programme from ~/Downloads/paper.pdf

It builds, and tells you how long the programme is, how it is structured, what it thinks the key ideas are, and whether anything failed the design rules. Nothing is uploaded anywhere; the files land in ~/mimem/programmes/.

Show me the figures

If the paper is a PDF, mimem has cropped each figure out of the page. next_figure hands one over as an image, you describe it, and the description is checked before it is spoken — the same gate the glosses face, plus one rule of its own: a description may not quote a value the paper never wrote in a sentence. Numbers read off a plot cannot be checked against anything, so they are rejected whether they are right or wrong. "Well over half", not "sixty point two seven percent". The exact values stay in study.md, next to the picture, where you can check them.

That rule is not caution. Given a real four-panel figure, a vision model returned a fluent description with two gases swapped and a 60.27% label read as "~69%" — so no retrieval card ever takes its answer from a figure description either. One wrong spoken sentence is a cost worth the feature; the same sentence drilled at expanding intervals is not.

Now write the explanations yourself

This is the good part. The assistant fetches the concepts that want explaining, along with the paper's own sentences about each one, writes the glosses, the concrete images and the analogies, and sends them back.

And they are checked. Every number in what it wrote is looked up in the sentences it was given; a claim that reverses a direction the paper states is rejected. Not warned about — rejected, and reported back so it can try again. This is the same check that runs when a paid API writes them, which is why letting the conversation do it is safe:

Rejected — gloss for "solid electrolyte interphase": number '41.85' does not appear in the cited source. Rewrite it using only what the source sentences say.

Read me the opening

One section at a time. A hundred-minute programme is fifteen thousand words, so the tools return summaries and let you ask for the parts you want.

The part that makes it worth doing¶

Quiz me on what's due

Every question mimem generated is a card, and answering one is recorded. Get it right and it comes back later; get it wrong and it comes back tomorrow. That is the spacing effect, which is the largest thing in the science and the reason any of this exists.

Ask for a session the next day and it will know which cards are ready and which you have never seen.

The log is a plain text file at ~/mimem/reviews.jsonl, one line per answer. You can read it, graph it, or delete it.

Where things are¶

audio.md is what you hand to a speech engine. mimem does not synthesise it yet — that is the next milestone — so for now: say -f audio.md on macOS, or paste it into a web reader.

What it can do¶

If it does not show up¶