Skip to content

mimem knowledge base — how humans learn from text, and what that means for audio

This knowledge base exists for one purpose: to constrain the design of the text that mimem generates. It is not a literature review for its own sake. Every entry ends with an implication that is turned into a numbered, testable rule in docs/DESIGN-RULES.md.

The problem we are designing against

Reading with your eyes is an actively self-regulated process. The reader controls the rate, re-reads a hard clause, jumps back three paragraphs to recover a definition, pauses to build a mental picture, and skips the parts they already know. Listening removes almost all of that control. Speech is transient: each word is gone the moment it is spoken, and the listener cannot cheaply go back.

So a naive text-to-speech reading of a book or paper strips out the machinery that made reading work, and adds nothing back. That is the whole reason the experience of "I listened to it and remember nothing" is normal rather than a personal failing.

mimem's thesis: if the listener can no longer regulate their own encoding, then the text itself must perform that regulation on their behalf — slowing down on hard concepts, re-stating, prompting imagery, and forcing retrieval — before it ever reaches the TTS engine.

Files

File Covers
01-core-effects.md Retrieval practice, spacing, generation, desirable difficulties, interleaving
02-audio-constraints.md Transiency, working memory, listening vs. reading, speech rate, segmentation
03-elaboration-and-imagery.md Dual coding, concreteness, elaborative interrogation, analogy, narrative, voice
04-structure-and-attention.md Advance organizers, pre-training, prequestions, signaling, coherence
05-verbalizing-nonprose.md Numbers, math, tables, figures, citations, code — the accessibility literature
99-sources.md Annotated source list with links

Reading the confidence tags

Effect sizes in education research are noisy, moderator-heavy, and subject to publication bias. Each finding carries a tag so that we know how hard to lean on it:

  • [A] Robust — large meta-analytic base, replicates across materials, ages and settings. Design decisions may depend on it.
  • [B] Solid but bounded — real effect, but with known moderators or a narrower evidence base. Use it, but make it configurable and do not build the architecture around it.
  • [C] Plausible / mechanistic — theory-driven or thin empirical base, or an engineering convention borrowed from accessibility practice rather than an experiment. Use as a default, expect to tune it against our own recall data.

Effect sizes below are quoted as reported by the cited sources (Cohen's d or Hedges' g); roughly, 0.2 is small, 0.5 medium, 0.8 large.