Most products model capability as something a user banks: they completed it, so they have it. They do not have it. From the moment of acquisition it is decaying, on a curve that differs per item and per person, and a product whose value depends on retained capability needs a maintenance loop rather than only an acquisition funnel.
This skill owns one decision: when to bring something back in front of a user. It does not own how hard the next item should be for that user — choosing when an item returns and choosing how hard to make it are different models, and difficulty belongs to the difficulty-calibration skill.
This skill is a graph: scan the descriptions, follow only the [[wikilinks]] you need.
How memory actually behaves
- [[references/spacing-beats-massing-at-equal-cost]] — the same repetitions spread across days survive; massed into one sitting they buy a passing score and nothing else. Frequency is the parameter, not length.
- [[references/sleep-is-a-free-consolidation-pass-you-can-skip]] — traces migrate to long-term storage overnight, so a one-block schedule deletes passes the user's brain would have run for free.
- [[references/decay-is-graded-and-fine-grain-goes-first]] — specifics rot fast, structure is durable, recognition outlives production. Decay is a shape, not a switch.
Scheduling: what comes back, and when
- [[references/per-item-half-life-not-a-fixed-ladder]] — fit a decay model to your own logs and schedule where lag approximates half-life, instead of walking a ladder someone picked in a meeting.
- [[references/delay-the-retry-and-weight-the-queue]] — an immediately re-served item is answered from working memory; hold the retry, and order the queue on recency and error.
- [[references/own-errors-are-the-highest-value-practice-set]] — a session built from this person's recorded failures spends every slot at their frontier; a population-average set wastes most of them.
- [[references/make-forward-motion-do-the-reviewing]] — if step N+1 structurally requires step N, retrieval happens as a side effect of the thing the user already wants to do.
Designing the practice itself
- [[references/retrieval-not-re-exposure]] — review that shows the answer strengthens recognition; review that makes the user produce it cold strengthens what real use requires.
- [[references/three-conditions-for-practice-to-pay]] — target the unmastered, require active retrieval, return feedback immediately. Drop one and you are burning the user's time.
- [[references/vary-the-encoding-context]] — vary the surfaces, voices and settings aggressively, or the capability binds to the one context you trained it in.
Maintenance, lapse and comeback
- [[references/give-mastered-things-a-visible-decay-state]] — a permanently-complete badge is a lie; render the decay on the object so the review request lives in the map instead of a notification.
- [[references/what-protects-a-skill-through-a-lapse]] — depth reached before the gap, a real reason to use it, a community and positive feeling predict survival. Hours logged does not.
- [[references/reactivate-dont-restart]] — forgotten capability is inaccessible, not destroyed. Probe for the residue; never hand a returner the beginner flow.
Sibling skills
- [[../duo-retention/SKILL]] — habit loops, streaks and churn diagnostics: how users come back at all, before this skill decides what to show them.
- [[../duo-gamification/SKILL]] — progression, XP and anti-grind: the systems most likely to quietly delete your maintenance loop in a redesign.
- [[../duo-experimentation/SKILL]] — how to run the test that spacing will initially appear to fail.
- [[../duo-product/SKILL]] — taking the long view, which is the only frame in which a decay model is worth funding.
Sources
All claims trace to dated posts on blog.duolingo.com, cited inline in each node's ## What Duolingo does section; slugs are verified against scripts/sources.json.