1. Anki's SM-2 Algorithm: Intervals and Memory Decay

Anki uses a modified SuperMemo-2 (SM-2) algorithm to schedule reviews. In SM-2, each card has an easiness factor (EF, default 2.5) and an interval (in days). For the first two successful reviews, Anki sets fixed intervals (typically 1 day then 6 days). Thereafter, if you answer correctly (quality $text{\geq}$ 3 on a 0–5 scale), the new interval is calculated as Interval_n-1 × EF¹ (rounded up), and EF is adjusted by the formula $EF' = EF +$ $(0.1 - (5-\text{quality})(0.08+(5-\text{quality})(0.02))$ , with a minimum EF of 1.3¹. If you fail (quality < 3), the card "resets" to the first learning step (interval = 1 day) but EF remains unchanged¹. This pair of rules – an exponential growth of intervals modulated by EF, and resetting on failure – models the forgetting curve: correct recall increases the card's "stability" (making EF larger), while failure forces a shorter interval to counteract memory decay. The effect is that easy cards quickly reach very long intervals, whereas difficult or forgotten cards are reviewed much sooner.

Anki's implementation makes some practical tweaks to classical SM-2². For example, Anki lets you customize the initial learning steps instead of hardcoding 1d and 6d, and it uses four response options ("Again", "Hard", "Good", "Easy") rather than SM-2's six-point scale². Negative answers ("Again") send the card into relearning (interval $\rightarrow$ 1) but – in learning mode – do not penalize the EF $\left( \begin{array}{c} 1 \\ 0 \end{array} \right)$. "Hard" answers reduce EF by 0.15 (and set a modest next-interval bonus), whereas "Easy" increases EF by 0.15 plus an extra interval multiplier 9. Anki also floors EF at 1.3 (130%) so that intervals never shrink too much². Late reviews (studying a card after its due date) are compensated by adding the delay into the next interval calculation².

These details ensure Anki's SM-2 tends to lengthen intervals for cards you recall easily (accelerating forgetting decay), and shorten or reset intervals for ones you find hard. In other words, memory decay is modeled by the ease factor: a higher EF means slower decay (longer intervals)¹. Over time, Anki's scheduler thus approximates each item's forgetting curve: by tracking performance (quality ratings) it updates EF to reflect how quickly you forget that item. This simple model is computationally light and has been shown to be effective for many learners, but it also has limitations (e.g. it treats each card independently and cannot directly use patterns from related cards).

2. Modern SRS Algorithms: FSRS, Ebisu, and Beyond