SM-2 — the algorithm behind Anki and most spaced repetition tools — was published in 1987 and is remarkably durable. It is also wrong in specific, predictable ways that matter once your deck grows.
What SM-2 does
Each card carries an ease factor, starting at 2.5. Grade a review 0-5; on success the interval multiplies by ease, on failure it resets:
if quality >= 3:
interval = 1 if reps == 0 else 6 if reps == 1 else round(interval * ease)
reps += 1
else:
reps, interval = 0, 1
ease = max(1.3, ease + (0.1 - (5-quality) * (0.08 + (5-quality) * 0.02)))
Simple, no training data, works offline. Genuinely good engineering for 1987.
Failure 1: ease hell
Every lapse decrements ease and it recovers slowly. A card you failed a few times early gets pinned near the 1.3 floor — permanently reviewed at short intervals even after you have learned it perfectly.
The algorithm has no way to say "this card was hard then, it is easy now." Ease is a one-way ratchet in practice.
Failure 2: binary-ish grading carries too much weight
The 0-5 scale collapses in practice to "got it / did not." But how you recalled it matters — instant recall and dredging it up after eight seconds are very different memory states that produce identical grades.
Response latency is a strong signal and SM-2 ignores it entirely.
Failure 3: cards are treated as independent
They are not. Learning "kanji A" helps with "compound containing A." Confusable pairs interfere with each other. SM-2 schedules every card in isolation, so interference-prone cards get scheduled adjacently by chance, which is the worst possible arrangement.
Failure 4: no forgetting curve per user
SM-2 assumes one exponential decay shape. Real retention varies by person, by material type, and by time of day. A fixed curve over-schedules some users and under-schedules others, and neither group can tell.
What modern approaches change
FSRS models memory with three variables — difficulty, stability, retrievability — and fits parameters to your review history. Concretely that means:
- Difficulty is not a one-way ratchet
- You can target an explicit retention rate (say 0.9) rather than accepting whatever falls out
- Latency and grade both inform the update
The cost is that it needs review history to fit, so cold-start is worse. Reasonable compromise: SM-2 defaults until you have a few hundred reviews, then refit.
I work on this at SmartRecall — AI-generated cards with scheduling that does not pin everything to the ease floor.
Caveat
For small decks with few lapses, SM-2 is fine and the added complexity buys nothing. These failures compound with deck size and time — they are invisible at 200 cards and painful at 5000.
Top comments (0)