Anki uses machine learning to choose when your flashcards return

Anki can use the history of your answers to decide when a flashcard should return. Its Free Spaced Repetition Scheduler, or FSRS, fits a memory model to past reviews. The machine-learning task is to estimate future recall and turn that estimate into a revision schedule.[1]

The educational routine remains familiar: see a question, try to retrieve the answer, then check it. Anki supports text, pictures, audio and other material, so the same routine can serve vocabulary, geographical knowledge or an illustrated anatomy question. Spacing those attempts across time gives a growing collection of cards a practical review sequence.[4]

Three estimates behind a due date

FSRS represents each card with three quantities. These describe its estimated memory state for that learner; the difficulty of one card need not match another card in the same deck.[2]

For an illustrative card with a stability of 20 days, the model expects recall probability to reach 90% after that interval. Another card can have a different stability. This is how the same learner can receive different schedules for material studied on the same day.[2]

The project reports that FSRS can require 20–30% fewer reviews than Anki’s legacy SM-2 scheduler at comparable retention. Its documentation identifies this as a simulation result, so the percentage describes modelled workload.[2]

Your answer supplies the feedback

After showing an answer, Anki asks the learner to judge the attempt. This creates the observations from which the scheduler learns. A correct answer recalled with difficulty and an answer that could not be recalled convey different information, even if both felt challenging.[3]

The displayed buttons also show when the card would return. Anki applies a small random adjustment to review intervals so cards learned together do not remain permanently grouped.[3]

If a review is late, the later answer helps the scheduler account for the actual delay.[2]

Choosing a realistic revision workload

Desired retention controls the target probability of remembering a card when it is due. The default is 90%. Raising it produces more frequent reviews, with workload increasing sharply near 100%. The separate Optimize command fits parameters to your history; the documentation recommends doing this about monthly.[1]

Anki’s experimental Help Me Decide feature simulates the workload associated with different retention targets. It offers a concrete way to compare a revision ambition with the minutes or reviews available each day.[1]

Consider a learner preparing a collection of historical dates alongside questions about the causes of an event. Remembering the date and explaining the causal argument are different tasks. The prompts determine which task is practised; the scheduler determines when that prompt returns.

How the forecasts are checked

The project’s public benchmark compares predicted recall with later review outcomes. Its evaluation excluding same-day reviews contains 349,923,850 reviews from 9,999 Anki collections. FSRS-6 records a mean log loss of 0.3460, compared with 0.3945 for a constant-average baseline. Lower log loss means the probability forecasts fit the observed outcomes better.[5]

That evaluates prediction on recorded reviews. The benchmark includes other memory models and neural networks with different inputs, and provides separate results when same-day reviews are included. These scores measure the quality of a memory forecast; effects on examination grades would require a separate educational evaluation.[5]

Making a collection useful over time

A useful flashcard asks for something the learner needs to retrieve. Trying the answer before revealing it makes that retrieval visible; a failed attempt identifies material to revisit. The benefit of a scheduled collection is continuity: familiar knowledge and difficult material can both remain part of an ongoing study routine.[4]

Report an error or suggest a correction