01 — Core memory effects¶

The four levers that actually move long-term retention, ordered by how much weight mimem puts on them.

1.1 Retrieval practice (the testing effect) — [A] Robust¶

Trying to produce information from memory is a far stronger memory event than re-exposing yourself to it. Meta-analyses report a medium effect of testing over restudy (g ≈ 0.50), holding up in real classrooms (g ≈ 0.50) and transferring to new questions about the same material (d ≈ 0.40).

Two moderators matter for us:

Response congruency. The retrieval prompt should require the same kind of operation the final use requires. If the goal is to explain a mechanism, the prompt must ask for the mechanism, not for a term.
Elaborated retrieval / feedback. Retrieval attempts followed by the correct answer are reliably better than unanswered attempts, and elaborated retrieval (retrieve and explain) outperforms bare recall. Feedback also rescues failed retrievals, which is essential in audio where we cannot check whether the listener answered.

Implication for mimem. The single highest-value thing the generated document can contain is question → silence → answer structures, placed inside the narration itself. Every retrieval beat must be followed by the answer, because a listener who fails silently gets nothing otherwise. → Rules RET-*.

1.2 Spacing and spaced retrieval — [A] Robust¶

Repetitions separated in time beat repetitions massed together. Cepeda et al.'s meta-analysis covered 254 studies / >14 000 participants and found distributed practice superior in 259 of 271 comparisons. Combining spacing with retrieval (spaced retrieval practice) gives a strong benefit over massed retrieval (g ≈ 0.74, Latimier et al. 2021).

Two useful details:

Gap size scales with the target retention interval. The common rule of thumb is that the optimal inter-study gap is roughly 10–20 % of how long you want to remember it. Want it in a month → review every 3–6 days.
Expanding vs. uniform intervals barely differ (g ≈ 0.03) unless there are many repetitions, where expanding starts to win. So: don't over-engineer the schedule.

Implication for mimem. Repetition inside a single document must be spaced, not adjacent. A key claim gets restated at the end of its own section, again at the end of the chapter, and again in the closing review — at deliberately increasing distances, measured in estimated listening minutes, not in paragraphs. Cross-document/cross-session spacing is Part 2, but Part 1 must already emit the per-item exposure record that Part 2 will schedule from. → Rules SPC-*.

1.3 Generation effect and desirable difficulties — [B] Solid but bounded¶

Information you produce yourself is remembered better than information handed to you. This sits in the broader family of desirable difficulties: conditions that feel harder and slower during study but produce better retention. The mirror image is the fluency illusion — smooth, easy-feeling exposure (re-reading, listening passively to a well-produced narration) inflates confidence without improving memory, and makes learners stop studying too early.

Important boundary condition: for high element-interactivity material (many interacting components that must be held in mind at once, e.g. a derivation), added difficulty becomes undesirable — it consumes the same working memory the learning needs. Also, a 2025 conceptual replication found text-generation benefits for expository text are weaker than the classic word-list literature suggests.

Implication for mimem. Insert difficulty deliberately and sparingly, and never in the middle of a dense derivation. Prefer generation prompts (predict, explain, produce an example) on consolidated material, and prefer support (worked-out narration, glosses, chunking) on dense material. The generated document must therefore know how dense each passage is. → Rules GEN-*, DIF-*.

1.4 Interleaving — **[B] Solid, but mostly not for us**¶

Brunmair & Richter's meta-analysis (59 studies, 238 effect sizes) found an overall benefit of interleaved over blocked practice (g = 0.42) — but it is highly moderated by between-category similarity. It was strongest for perceptual categories such as painters' styles (g = 0.67) and not significant for expository texts (g = 0.21) and actually negative for word learning (g = −0.39).

Implication for mimem. This is a case where the literature tells us not to build something. Do not shuffle expository content to interleave topics — our material is exactly the type where interleaving fails. Keep exposition blocked and coherent. Interleaving is applied only inside review/quiz blocks, where items from different sections are mixed, which is closer to the condition where it works and is anyway a spacing manipulation. → Rule SPC-04.

1.5 What does not work (so we must not accidentally build it)¶

Re-reading / re-listening as the primary strategy: low utility, high fluency illusion. Simply generating a "nicer" audio version of the same text and playing it twice is the failure mode we are trying to escape.
Highlighting and summarizing were rated low utility by Dunlosky et al. (2013) for typical learners — the value is in generating the summary under retrieval conditions, not in owning one. An auto-generated summary that the listener passively consumes buys much less than it appears to.
Massed repetition ("say it three times in a row"): near-zero long-term benefit compared to the same three repetitions spaced out.

Implication. Every repetition mimem emits must be spaced, varied in form (paraphrase, not verbatim), and where possible converted into a retrieval attempt rather than a restatement. → Rules REP-*.