Tuning it to you¶

Two files decide what the programme is like. One says how hard it should work you; the other says what you already know. The second matters more.

listener.yaml — what you already know¶

This is the single biggest quality lever for anyone who has a field. It is what stops the system explaining "electrolyte" to you for the hundredth time.

name: jepe
default_expertise: familiar        # novice | familiar | expert

expertise:
  electrochemistry: expert
  machine-learning: familiar

# What vocabulary belongs to each domain. Without this, declaring expertise does nothing:
# no concept in a real paper is called "electrochemistry".
domain_terms:
  electrochemistry: [anode, cathode, electrolyte, cell, electrode, lithiation]
  machine-learning: [model, training, validation, hyperparameter]

# Terms you know exactly. Knowing "anode" also makes "anode active material" easier.
known_terms: [coulombic efficiency, C-rate, SEI, state of charge]

# How to say things a text-to-speech engine will get wrong.
lexicon:
  NMC811: "N M C eight one one"
  LiFePO4: "lithium iron phosphate"
  Li-ion: "lithium ion"

What it actually changes¶

On a battery teardown paper, for a declared electrochemistry expert:

The programme does not just skip a few definitions. Where the effort goes changes, so the subject of the programme changes: for a specialist, the methodology; for a newcomer, the objects.

profiles/*.yaml — how hard it works you¶

Three ship with mimem. Pick with --profile.

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

Every knob in those files is documented in place, and each names the design rule it enforces. The ones people change most:

wpm

Speech rate for the duration estimate. Default 155, which is a normal technical narration pace. Comprehension holds up to about 270 and falls off a cliff after — the config refuses anything outside 80–270.

numeric_fidelity

exact by default. Every value is spoken at the precision the paper states. Set to rounded for two significant figures in running prose, which never applies to a value a retained claim turns on.

spacing.min_gap_minutes

How far apart two encounters with the same idea must be. Default 3. Lower it and you get more repetition that does less good; the whole point of spacing is that the forgetting between exposures is what makes the next one work.

max_preload_terms

How many terms are defined before the paper starts. Default 7, and the number is not arbitrary — more than a handful of new labels at once is exactly the load that pre-training exists to avoid.

The elaboration layer¶

Off by default. Switched on, a model writes the glosses, the concrete images, the why-explanations and the analogies — the parts that are genuinely writing problems rather than structural ones.

uv run mimem elaborate paper.ir.json --dry-run

29 call(s) to claude-opus-5: 3 x analogy, 4 x anchor, 11 x gloss, 11 x why
  estimated $1.67 -- an estimate, not a quote

Then, if that looks reasonable:

uv run mimem build paper.pdf --llm --budget 3.00 --out out/paper

--budget is a hard cap, checked before each call rather than reported after it. Answers are cached, so re-running a build after editing one paragraph does not re-pay for the document.

Everything else¶

Both files are plain YAML with comments, and mimem profiles prints what is currently loaded:

uv run mimem profiles

If you want to know why a knob exists at all, every one of them names a rule in the design rules, and every rule names the finding behind it in the knowledge base.