When will you forget this?

Flashcard apps schedule reviews with a fixed rule: get it right, wait twice as long. But how fast a thing fades depends on the thing and on you. This model learned that decay from 10.2 million real reviews, and was then tested on 1.3 million more from learners it had never seen. It schedules from what it learned.

Loading the model and held-out results…

Something you are learning

Times you recalled it — 4

Times you forgot it — 1

learned decay curve when Leitner would review you when the model would review you (90% recall)

Learned half-life

—

the point where you are as likely to forget it as recall it

Leitner would wait

—

Put a word into practice

Add an answer, try recalling it, then record how it went. Your own review counts are separate from the sliders above.

Opening local study storage…

Model times are estimates from other learners' review logs, not proven recommendations. Your cards and review history stay on this device; clearing browser data removes them. Nothing you enter is uploaded.

Does the learned schedule actually predict better?

Held out by learner — no person's reviews appear in both training and test data, so nothing is scored on someone the model has already met. Lower error is better.

These are predictions from observational review logs, not evidence that following this schedule improves learning. The original app chose when to test each learner. Its scheduling policy can affect the fitted weights, including the small negative weight on prior successful recalls. The model has lower mean absolute error here; Leitner has higher AUC. Neither comparison proves that the displayed interval is best for an individual learner.