Forgetting is not the enemy of learning.

It is the clock that makes scheduling necessary.

If every memory stayed sharp forever, studying would be a packing problem: put facts into the head once, move on. If every memory vanished instantly, studying would be hopeless. Human memory lives in the interesting middle. Traces decay, practice changes the decay rate, retrieval can strengthen the trace, and the right next review depends on the future date you care about.

That turns learning into a scheduling problem.

The naive question is:

How often should I review this?

The better question is:

When will predicted recall fall below the level I am willing to risk?

A flashcard app is not just a pile of cards. It is a tiny control system with a daily budget, trying to keep thousands of memories above a threshold.

The Curve Does Not Pick the Card

Ebbinghaus made forgetting measurable. In Memory: A Contribution to Experimental Psychology, he used himself as the subject, studied nonsense syllables, and measured retention and relearning across delays.1 The famous curve is easy to over-simplify, but its central lesson is still alive: retention changes quickly at first and slowly later.

That fact alone does not tell you when to review.

Suppose recall follows a simple half-life model:

P(recall after t days) = 2^(-t / h)

where h is the current half-life of the item. If you want recall to stay near target, the next review should arrive around:

t_due = h * log(target) / log(0.5)

For target = 0.8, that is about 0.32 h. For target = 0.5, it is h. The same memory can be reviewed soon or late depending on the operational standard. A medical fact, a language flashcard, and a theorem lemma do not need the same risk budget.

This is why “review every day” is wasteful and “wait until you forget” is too late. The useful zone is between boredom and collapse.

The Ridge, Not the Rule

The spacing effect is robust, but it is not a single magic interval.

Cepeda, Pashler, Vul, Wixted, and Rohrer reviewed 839 assessments of distributed practice across 317 experiments in 184 articles. Their synthesis found that the best interstudy interval depends jointly on the delay to the final test.2 A later experiment made the ridge visible: for a given test delay, increasing the gap first helped and then hurt. The optimal gap grew with test delay, while the optimal gap as a fraction of test delay shrank from roughly 20-40 percent for a 1-week delay to roughly 5-10 percent for a 1-year delay.3

That is the sentence most school calendars ignore.

Spacing is not simply “longer is better.” If the gap is too short, retrieval is too easy and the practice does little. If the gap is too long, the learner fails or relearns from scratch. The optimum is a ridge, not a point.

Pavlik and Anderson framed practice scheduling as a model-based balancing act: spacing, recency, and frequency pull in different directions, and an optimized condition can outperform hand-designed schedules.4 Settles and Meeder later brought this into a production language-learning setting with half-life regression. Their Duolingo model estimated item half-lives from practice history, reduced recall-rate prediction error by more than 45 percent relative to baselines, and improved daily engagement by 12 percent in an operational study.5

The modern version is even more explicit. Tabibian and coauthors model spaced repetition as an optimal-control problem: review schedules are chosen to maximize retention while controlling review effort.6 That is exactly what a review queue feels like when the deck is large. The hard part is not knowing that spacing helps. The hard part is deciding which item gets the next scarce review.

Queue Lab

The lab below simulates a small deck of memories. Each item has a difficulty and a current half-life. Recall probability decays as:

P = 2^(-days_since_review / half_life)

When an item is reviewed, the model samples a deterministic recall outcome from that probability, gives feedback, and updates the half-life. Reviews that happen when recall is already very high are counted as too easy. Reviews that happen after recall has fallen too low are counted as too late.

Four scheduling policies share the same deck, horizon, new-card rate, and daily review budget:

  • Massed spends early reviews quickly and then leaves items alone.
  • Fixed reviews every item at a fixed cadence.
  • Leitner moves cards through boxes with longer intervals after successful recall and shorter intervals after failure.
  • Half-life reviews the most at-risk due items when predicted recall falls below the selected target.
Massed Fixed Leitner Half-life Review debt

Deterministic synthetic model. The point is not that these parameters are universal; the point is that a scheduler must trade retention against review effort, and the right interval depends on the target horizon and item difficulty.

The default setting is intentionally crowded: 4 new cards per day, 18 reviews per day, and a 90-day horizon. The half-life controller usually keeps more of the deck above target because it spends reviews near the edge of forgetting instead of on cards that are already easy.

Raise Target recall to 90%. The half-life policy will review earlier and more often. Lower the Review budget to 10 / day; the queue starts to carry debt. The scheduler is not magic. It can only allocate scarce retrievals.

Switch Highlight to Massed. Its early recall looks fine because it spends reviews immediately. The final exam-day recall is the warning: practice that is optimized for today’s fluency can be badly scheduled for future access.

Memory Budget Interface

The next useful learning system would not expose a single “due” date. It would expose a memory budget:

  • predicted recall now;
  • predicted recall at the exam or use date;
  • half-life estimate and uncertainty;
  • review cost for moving the item above target;
  • opportunity cost of reviewing this item instead of another;
  • queue debt created by adding new material today.

That interface would make the learner’s real trade-off visible. New cards are not free. A daily streak is not a retention guarantee. Review ease is not the same as review value.

The best review is often the one that feels slightly uncomfortable and arrives just before the memory would become expensive to recover.

That is the strange kindness of forgetting: it tells the scheduler where effort has become useful again.

  1. Hermann Ebbinghaus, Memory: A Contribution to Experimental Psychology, 1885, translated by Henry A. Ruger and Clara E. Bussenius, 1913. Classics in the History of Psychology, PMC edition. ↩

  2. Nicholas J. Cepeda, Harold Pashler, Edward Vul, John T. Wixted, and Doug Rohrer, “Distributed Practice in Verbal Recall Tasks: A Review and Quantitative Synthesis,” Psychological Bulletin, 2006. PubMed, PDF. ↩

  3. Nicholas J. Cepeda, Edward Vul, Doug Rohrer, John T. Wixted, and Harold Pashler, “Spacing Effects in Learning: A Temporal Ridgeline of Optimal Retention,” Psychological Science, 2008. SAGE abstract, PDF. ↩

  4. Philip I. Pavlik Jr. and John R. Anderson, “Using a Model to Compute the Optimal Schedule of Practice,” Journal of Experimental Psychology: Applied, 2008. University page. ↩

  5. Burr Settles and Brendan Meeder, “A Trainable Spaced Repetition Model for Language Learning,” ACL, 2016. ACL Anthology, PDF, code and data. ↩

  6. Behzad Tabibian, Utkarsh Upadhyay, Abir De, Ali Zarezade, Bernhard Scholkopf, and Manuel Gomez-Rodriguez, “Enhancing Human Learning via Spaced Repetition Optimization,” PNAS, 2019. PNAS, arXiv version. ↩