Skip to content

Without a key

This page is first because it covers most people. You need no API key, no account and no local model to get everything mimem is actually for.

What works with nothing configured

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

That is the whole pipeline: the paper read, triaged, verbalized, laid out as a programme with questions and pauses and spacing, rendered to four files, and checked against thirty lint rules. It exits non-zero if it broke one.

You get audio.md, study.md, cards.json and manifest.json. Add --speak sapi on Windows, or --speak silent anywhere, and you get audio.wav too — see Turning it into audio.

What you do not get is the explaining layer: no one-line glosses beyond the ones the paper writes for itself, no concrete anchors, no analogies, no figure descriptions. The manifest lists each one as degraded, with the fallback it took. Nothing is invented to fill the gap.

The assistant route: a model, still no key

The best option for most people. mimem ships an MCP server, so the assistant you are already talking to does the explaining — no key, no billing, and you watch every number get checked.

Download mimem.mcpb, 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". Full instructions, including the settings the extension exposes, are on Using it from Claude or ChatGPT.

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

Create .cursor/mcp.json in your project, or ~/.cursor/mcp.json for every project:

{
  "mcpServers": {
    "mimem": {
      "command": "uvx",
      "args": ["--from", "git+https://github.com/jepegit/mimem",
               "--with", "mcp", "mimem-mcp"],
      "env": {"MIMEM_WORKSPACE": "~/mimem"}
    }
  }
}

Settings → Connectors → Advanced → Developer mode, then add a connector of type STDIO. The command is the same uvx line as above.

All of these need uv on your PATH; mimem doctor checks for it.

Which of these are verified

The Claude Desktop extension is tested end to end — installed, eleven tools listed, tools called. The underlying server is verified to speak MCP correctly over stdio, which is what every host on this page uses. The specific config files for Cursor and ChatGPT follow each host's documented format and have not been tested here; if one is wrong, please open an issue and it will be fixed rather than left to be discovered again.

What the assistant can and cannot do

It writes the explanations, describes figures, and runs study sessions across days. It cannot synthesise audio — speech is a CLI command, because a speech engine is a different kind of thing from a conversation. Build the programme through the assistant, then:

uv run mimem speak ~/mimem/programmes/<id> --engine sapi

Recording a fixture set

If you ever do run a paid model, keep the answers:

uv run mimem record paper.pdf --out fixtures/paper
uv run mimem build paper.pdf --fixtures fixtures/paper   # free, forever, identical

One paid run becomes a set anyone can replay for nothing. It is how the worked example in these docs stays reproducible and how the test suite stays free.