v0.1.0 — document in, script out

mimem rewrites a document so that the text itself does what a careful reader would have done — because listening removes every one of those moves, and a speech engine puts nothing back.

Read this first: there is no voice yet. mimem produces a script designed to be spoken, not a sound file. Handing that script to a speech engine is M7, the next milestone. If you came here for an audio file in one command, this release is early for you.

What it does produce is a programme: an orientation, a few questions to hold on to, terms defined before they are used, the paper in segments that each end on a question with a pause long enough to actually try it, and the ideas that matter coming back at growing intervals.

Two ways in

Claude Desktop. Download mimem.mcpb below, then Settings → Extensions → Install extension and pick the file. Then say "make me a study programme from this paper", and days later, "quiz me on what's due". The assistant writes the explanations, so no API key is needed. Setup guide →

A terminal.

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

Four files come out: audio.md (what the engine says — no digits, no brackets, no citations), study.md (the written companion, with the source sentence and page behind every beat), cards.json, and manifest.json (the audit trail).

Both need Python 3.11+ and uv. The extension is a manifest and an icon — uv fetches and builds the server on your own machine at first start, so there is no platform matrix to pick from.

What is in it

Nine stages from PDF to script: ingest, clean, triage, concepts, verbalize, elaborate, plan, render, lint. Concepts scored by difficulty × importance and shifted by what you already know. A spacing scheduler. Ninety design rules checked on every build, with a non-zero exit, so a script that breaks one cannot be produced quietly.

Nothing is invented. Every sentence is either the paper's own — carrying the exact place it came from — or one of about a dozen named templates. Where a model does write an explanation, every number in it is checked against the sentences it was written from, and a claim that reverses the paper's direction is rejected. Figures are described by a model that can see them, under the same rule: it may not state a value the paper's own text does not, and it is refused when it tries. That check exists because a real vision model, on a real figure, swapped two labelled species and misread 60.27% as roughly 69% — fluently, and with no sign anything was wrong.

Known limits

• No speech synthesis. M7.
• The live API adapter has never been run. It was written in an environment with no key. The Claude Desktop route does not use it — the assistant writes the explanations in the conversation — so the tested path and the untested one are different paths.
• Evaluation is mechanical only. It measures whether the output still has the properties the design asks for. It does not measure whether anyone learned anything; that needs a model, and then people.
• Eight sentences in a 98-page test paper still open in lower case, all of them list enumerations or figure panel letters. Known, documented, deliberately not papered over.

Verified for this release

The bundle's launch command was run against GitHub rather than a fixture: uvx --from git+…@v0.1.0 --with mcp mimem-mcp installs, launches, speaks MCP, and answers with its eleven tools. The --from ref is pinned to this tag, so this download installs this code and not whatever main becomes.

MIT. Full design and reasoning in the plan and the rules.