Using a model with mimem¶

None of this is required. mimem's whole argument is that structure matters more than fluency, and the structure is deterministic: a document goes in and a listenable, checkable programme comes out with no model involved at all. What a model adds is the explaining — a one-line gloss, a concrete anchor, an analogy that says where it breaks down, a description of a figure you cannot see.

Start here:

uv run mimem doctor

It says what is reachable from your machine and, for everything that is not, the exact next thing to type. On a machine with nothing configured it looks like this — which is a real run, on the machine mimem is developed on:

text generation
  assistant  ready         no key needed; the model is the conversation
  anthropic  no key
             -> export ANTHROPIC_API_KEY=sk-ant-...
  openai     no key
             -> export OPENAI_API_KEY=sk-...
  local      missing       nothing listening on 11434, 1234, 8080, 8000
             -> ollama serve   (or start LM Studio / llama-server / vLLM)

speech
  silent     ready         always available; makes shaped silence
  sapi       ready         9 voices installed
  piper      missing       not on PATH
  openai     no key

tools
  ffmpeg     ready         available for converting audio.wav
  uv         ready         the extension needs it at first start

mimem doctor --live sends one tiny structured request to each configured provider, because the key is set and the key works are different claims.

Four ways in, in the order most people should try them¶

The assistant route is first on purpose. It inverts the problem: the model is the conversation you are already in, so there is no key, no billing and no provider to configure. It is not a workaround for people who cannot afford an API — for most users it is simply the best option.

What a model is and is not allowed to do¶

The division is the whole safety argument, and it is worth knowing before you turn any of this on:

Deterministic code owns every fact. Numbers, units, symbols, what gets dropped, what comes back and when — none of that is ever a model's decision. Rule NUM-01 and the verbalizers own values; the planner owns structure.

A model owns prose, and every sentence it writes is checked. Every number in a generated sentence must appear in the source sentences it was written from; a claim that reverses the paper's direction is rejected before it reaches the programme (GRD-03). You can watch that happen — the elaboration report lists what was accepted, what was rejected and why.

When the model is absent, every task degrades along a documented path and the manifest records why — there are four different reasons and they need four different actions from you: no provider configured, configured but unreachable, reached but refused the shape, or out of budget. A run with no provider at all says so once rather than once per task. A gloss falls back to the paper's own definitional sentence; an anchor is omitted rather than invented; a figure falls back to its caption. That is why --local is the default and why a forgotten flag cannot start billing you.

Was it worth it?¶

uv run mimem compare paper.pdf --fixtures fixtures/paper

Builds the document twice — once deterministic, once with the model — and diffs the result. The deterministic build is the control; the difference is what you paid for.

gloss_coverage      0.182 -> 0.273  (+0.091 better)
concepts_spaced     0.545 -> 0.364  (-0.181)
cards               10 -> 7         (-3)
prompts_per_minute  0.831 -> 0.697  (-0.134)
the model wrote: 1 analogy, 1 anchor, 1 gloss, 1 why
cost: $0.0000 over 0 calls

That shape is worth understanding before you spend anything. More terms get defined — and the explaining costs duration, which comes out of the same budget as the questions. Fewer cards, fewer prompts per minute, ideas coming back less often. Retrieval practice is the second-largest effect in the knowledge base, so a build that explains more and asks less is not automatically a better one.

Whether that trade is right depends on the document and on you. The point of the command is that it is now a number you can look at rather than an impression.

Which implementation answers each task¶

Each task can be set to one of three modes in the profile:

Everything defaults to prefer. gloss looked like the obvious assist candidate — it is the one task with two real implementations — and that was tried and reverted, because the two do not produce the same thing: the rules write the pre-load's one line, the model writes that and the full introduction. Skipping the call because a short definition already existed would have given a paid run less for the same money, silently. It is offered, not assumed.

Cost, briefly¶

--dry-run prints the planned calls and an estimate without making any. --budget 2.00 stops the run rather than surprising you. Both work with every provider.

A model is called at most a few dozen times per paper — once per concept that earns an elaboration, once per figure, once per verification — and the source document is sent as a cached prefix, so the second task on a document costs a fraction of the first.