FSRS vs SM-2, explained

Where SM-2 comes from

SM-2 was published by Piotr Wozniak in 1987 for SuperMemo, and it earned its longevity. The idea that made it work is the one behind all spaced repetition: reviews spaced just before you forget beat reviews crammed together, and each successful recall lets the next interval stretch further.

Mechanically, SM-2 gives every card an ease factor that grows or shrinks with your answers, and multiplies the interval by it after each review. It is simple enough to compute on a 1980s PC, and it carried SuperMemo, Anki (for most of its life), and countless clones. But the simplicity has three costs:

• One curve fits all. The formula applies broadly the same schedule shape to every learner and every kind of material, whether you are drilling drug names or contract law doctrines.
• Lapses are brutal. Forget a mature card once and SM-2 knocks its progress back hard, even though forgetting one fact about a topic rarely means you lost the whole memory.
• It never estimates recall. SM-2 cannot tell you the probability that you still remember a card right now. It schedules blind, so it cannot trade workload against retention deliberately.

What FSRS does differently

FSRS, the Free Spaced Repetition Scheduler, is an open-source algorithm developed since 2022 as SM-2's successor. Instead of one ease multiplier, it maintains a small memory model for every card, built on three quantities:

• Difficulty: how hard this specific card is for you, learned from how you have answered it.
• Stability: how long the memory currently lasts, measured as the time for recall probability to fall to 90%.
• Retrievability: the model's estimate of the probability you would recall the card right now.

The parameters behind those quantities are fitted with machine learning on hundreds of millions of real reviews, and refitted to your own review history as it accumulates. Scheduling then becomes an explicit decision instead of a formula's side effect: you pick a desired retention (say, 90%), and FSRS brings each card back at the moment its predicted recall is about to drop below that line. Cards you find easy drift to long intervals quickly; cards that keep tripping you return sooner; a single lapse updates the model rather than resetting the card.

What the evidence shows

The strongest public evidence is the open-spaced-repetition benchmark, which evaluates scheduling algorithms against roughly 10,000 anonymized Anki collections, about 350 million real reviews in total. Each algorithm predicts the outcome of every review; predictions are scored with log loss, which punishes confident wrong answers. Three results matter for this comparison:

• FSRS predicts recall more accurately than SM-2 in about 99.6% of collections. Almost nobody's review history fits the 1987 formula better than the fitted model.
• In workload simulations on the same data, FSRS reaches the same retention with roughly 20 to 30% fewer reviews than SM-2, because it stops re-showing cards the model is confident you still know.
• The underlying memory model is published research: the DSR model behind FSRS appeared in a peer-reviewed KDD 2022 paper, not a marketing page.

What the evidence does not prove

Most tools quote the 20 to 30% figure as settled fact. We would rather tell you exactly how far the evidence goes:

• It is a simulation, not a trial. The benchmark replays and models real review logs. Nobody has run a large randomized controlled trial where students are assigned to FSRS or SM-2 and sit the same exam.
• SM-2 is playing an away game. SM-2 was never designed to output recall probabilities, so comparing prediction accuracy requires bolting extra formulas onto it. The comparison is fair in spirit, but it is a reconstruction.
• Your number will vary. The saving depends on your material, your retention target, and how consistently you review. Treat 20 to 30% as a well-supported estimate, not a promise.

None of this weakens the practical conclusion. Better predictions on 99.6% of real collections is about as one-sided as evidence gets in this field. It just means the right claim is "almost certainly fewer reviews for the same retention" rather than a precise percentage.

What it means for your daily reviews

For a medical student six months from Step 1 or a bar candidate holding thousands of cards, the differences compound:

• Lower daily load at the same retention. The reviews FSRS cuts are the ones you were going to get right anyway.
• Lapses stop being catastrophic. Missing one mature card adjusts its model instead of resetting months of progress, which also makes the daily queue less demoralizing.
• Workload becomes a dial. Desired retention is an explicit setting: 90% is a sensible default, and you can see what pushing to 95% would cost in extra reviews before you commit to it.

How Cadence Deck uses FSRS

Cadence Deck schedules every flashcard with FSRS from the first review: real per-card difficulty, stability, and retrievability, not a fixed-interval loop with a spaced-repetition label. Each card shows its predicted recall and when you last saw it, so the schedule is never a black box. You can see how the three study modes share one FSRS schedule on the features page, or compare the wider tool landscape in our honest Anki alternatives comparison.

Further reading

• A Stochastic Shortest Path Algorithm for Optimizing Spaced Repetition Scheduling (KDD 2022): the peer-reviewed research behind the FSRS memory model.
• The open-spaced-repetition benchmark: methodology, data, and results for FSRS against SM-2 and other schedulers.
• Improving Students' Learning With Effective Learning Techniques: the broader evidence that distributed practice is among the most effective study techniques known.