Spaced repetition: seeing each question again at the right moment

This page describes exactly which algorithm we use, the figures it relies on and what it cannot do yet.

Two quantities per question, and the time that passes

We use FSRS, an open-source, actively maintained spaced repetition scheduler that Anki has used since 2023. Our implementation is the official TypeScript port of the algorithm, with its twenty-one parameters, and not an in-house approximation.

FSRS does not keep a score, but two quantities for each candidate-question pair. Difficulty, on a scale of 1 to 10, says how much this question costs you, personally. Stability, expressed in days, says how long your memory holds before it drops to the target recall level, set at 90 per cent.

The time elapsed since you last saw the question is part of the calculation. Answering correctly the next day and answering correctly six months later do not say the same thing about your memory, and therefore do not produce the same next interval.

What the algorithm observes in order to rate you

Most spaced repetition tools ask the user to rate themselves after the fact: easy, medium, hard. In competition training, that extra question breaks the rhythm and rests on a judgement that is unreliable just after seeing the correct answer. We infer the rating from what you actually did.

Speed is measured as a share of the time allowed for the question, not in absolute seconds: our durations run from 30 to 150 seconds depending on the question, and the same duration does not mean the same thing for a numerical reasoning question and for a knowledge question.

• Wrong answer: the question comes back the very next day, whatever your speed was. Answering fast and wrong buys you nothing.
• Correct answer in less than one tenth of the time allowed: the answer is judged fluent and the interval lengthens markedly.
• Correct answer taking more than six tenths of the time allowed: the answer is judged laboured and the interval lengthens only slightly.
• Correct answer between the two: normal pace, standard progression.
• You changed your answer before confirming: the rating drops by one step. A hesitation that is resolved is still a hesitation.
• More than ten minutes on a question: the time is ignored instead of being counted as slowness. Beyond that point, the time reflects an interruption, not your certainty.

When a question comes back, in concrete terms

On a run of correct answers at the normal pace, a new question comes back after 3 days, then 14, then 57, then 196. Beyond that, the interval is capped at one year: preparing for a competition takes months, and a question scheduled beyond the date of the test is of no use to anyone.

A wrong answer brings the due date back to the next day and lowers stability, without erasing what you had built: the difficulty the question has acquired is kept, and recovery is simply slower than learning it the first time.

A question counts as acquired after three consecutive correct answers at the normal pace, which corresponds to a next due date about two months away. Three laboured correct answers are not enough: the stability reached stays below the threshold, and the question stays in revision.

• Question of the day: a revision that is due comes before anything else, the oldest due date first. With no revision due, you receive a new question. If the bank for your language is exhausted, you receive the oldest acquired question.
• Targeted training: about a quarter due revisions, a tenth questions you have already acquired (to keep them fresh), the rest new questions. You choose from 1 to 50 questions, 20 by default.
• The same question is never put back in front of you less than one hour after you answered it, even if the algorithm judges it due.
• The training selector shows, test by test, how many questions are left to discover and how many are due for revision.
• A daily reminder can tell you when revisions fall due. It can be switched off in your notification preferences.
• If you have entered the date of your competition, the interval is additionally capped at half of the time remaining: each question is therefore put back in front of you at least twice before the test, once at the halfway point and once in the final stretch.

What we cannot do yet

The speed thresholds are provisional. They were calibrated on the actual distribution of our users’ correct answers, but on a sample that is still small. They will be recalibrated when the volume allows it, and this page will be updated on the day they change.

The parameters of the algorithm are the FSRS default parameters, the same for everyone. FSRS can optimise itself on each user’s personal history; we do not do that yet, for want of a history large enough for the optimisation to be better than the default rather than worse.

Finally, no manual adjustment is offered: you can neither force a question to come back earlier, nor take it out of the queue. We chose to keep things simple, and we will reconsider if you tell us that you need these adjustments.