What spaced repetition algorithm does Anki use?

Anki’s algorithm is based on the SuperMemo 2 algorithm. For info on SM-2, please see http://www.supermemo.com/english/ol/sm2.htm

Anki’s algorithm differs from SM-2 in some respects. Notably:

  • SM-2 defines an initial interval of 1 day then 6 days. With Anki, you have full control over the length of the initial learning steps. Anki understands that it can be necessary to see a new card a number of times before you’re able to memorize it, and those initial "failures" don’t mean you need to be punished by being shown the failed card many times over the course of a few days. Performance during the learning stage does not reflect performance in the retaining stage.

  • Anki uses 4 choices for answering review cards, not 6. There is only one fail choice, not 3. The reason for this is that failure comprises a small amount of total reviews, and thus adjusting a card’s ease can be sufficiently done by simply varying the positive answers.

  • Answering cards later than scheduled will be factored into the next interval calculation, so you receive a boost to cards that you were late in answering but still remembered.

  • Like SM-2, Anki’s failure button resets the card interval by default. But the user can choose to have the card’s interval reduced instead of being reset completely. Also, you can elect to review failed mature cards on a different day, instead of the same day.

  • Remembered easily not only increments the ease factor, but adds an extra bonus to the current interval calculation. Thus, answering remembered easily is a little more aggressive than the standard SM-2 algorithm.

  • Successive failures while cards are in learning do not result in further decreases to the card’s ease. A common complaint with the standard SM-2 algorithm is that repeated failings of a card cause the card to get stuck in "low interval hell". In Anki, the initial acquisition process does not influence a card’s ease.

You can also check out sched.py and schedv2.py in Anki’s source code for the scheduling code. Here is a summary (see the deck options section of the manual for the options that are mentioned in italics).

Learning/Relearning Cards

If you press…​

  • Again
    Moves the card back to the first step setted in Learning/Relearning Steps.

  • Hard
    Repeats the current step after the first step, and is the average of Again and Good.

  • Good
    Moves the card to the next step. If the card was on the final step, the card is converted into a review card (it 'graduates').

  • Easy Immediately converts the card into a review card.

New cards have no ease, so no matter how many times you press 'Again' or 'Hard', the future ease factor of the card won't be affected. The same can be said about relearning cards: pressing 'Again' or 'Hard' won't have any effect over the card's ease.

Review Cards

Once a card is graduated, it gets an ease factor. By default is 2.5, but you can set another value using the Deck Options.

If you press…​

  • Again
    The card is placed into relearning mode, the ease is decreased by 20 percentage points (that is, 20 is subtracted from the ease value, which is in units of percentage points), and the current interval is multiplied by the value of new interval (this interval will be used when the card exits relearning mode).

  • Hard
    The card’s ease is decreased by 15 percentage points and the current interval is multiplied by the value of hard interval (1.2 by default)

  • Good
    The current interval is multiplied by the current ease. The ease is unchanged.

  • Easy
    The current interval is multiplied by the current ease times the easy bonus and the ease is increased by 15 percentage points.

For Hard, Good, and Easy, the next interval is additionally multiplied by the interval modifier. If the card is being reviewed late, additional days will be added to the current interval, as described in a previous FAQ.

Limitations

There are a few limitations on the scheduling values that cards can take. Eases will never be decreased below 130%; SuperMemo’s research has shown that eases below 130% tend to result in cards becoming due more often than is useful and annoying users. Intervals will never be increased beyond the value of maximum interval. Finally, all new intervals (except Again) will always be at least one day longer than the previous interval.

Why doesn’t Anki use SuperMemo’s latest algorithm?

The simple answer is that SuperMemo’s latest algorithm is proprietary, and requires licensing. As Anki is an open source application, it can only make use of algorithms that have been made freely available.

We’re inclined to believe SuperMemo when they say their newer algorithms are more efficient, but feel that to a certain extent, it is a case of diminishing returns. The gains achieved by moving from a traditional study routine to SM-2 are already great, and by sticking with an open algorithm, your learning data is not locked into a single product.

Ultimately it’s up to you to decide - if access to the latest and greatest scheduler is a higher priority than the things that Anki brings to the table, you may want to check out SuperMemo to see if it is a good fit for you.

What about SM-5?

Anki’s scheduler was originally based on SuperMemo's SM-5. Anki’s default of showing the next interval above each ease button revealed problems with the implementation - harder cards could end up with greater interval increases than easy ones, and the ease factors sometimes grew to the point where a single review could result in a 20-30x increase in interval.

An attempt was made at the time to correct this by smoothing the optimal factors matrix - applying a cap on the maximum factor and enforcing a minimum difference between adjacent ease factors. This addressed the above problems, but resulted in an optimal factors matrix that had very little room to move, and the conclusion drawn at the time was that SM-5 was not an improvement over SM-2.

While SM-5 clearly wasn’t working for Anki, in hindsight, it may not have been fair to assume the issues we encountered were due to fundamental problems with the algorithm. SuperMemo have subsequently stated that the description of the SM-5 algorithm listed on their website is incomplete, so it is possible the problems we encountered do not exist in SuperMemo’s proprietary implementation.