Skip to content

FSRS vs SM-2

For most Anki users, FSRS is the better choice: in a benchmark on Anki review logs, it estimated the probability of recall far more accurately than SM-2. SM-2 remains simpler and more predictable, and it needs no review history.

Updated 5 min read

In short

  • SM-2 multiplies each interval by an ease factor that goes up or down with your answers. FSRS estimates the probability of recall and schedules a review when it falls to the retention you chose, such as 90%.
  • On about 350 million Anki reviews, FSRS fitted users' review outcomes much better than SM-2 on every metric. The benchmark is run by the FSRS developers.
  • The FSRS developers report 20–30% fewer reviews for the same retention, but that figure comes from simulations, not from a controlled study.
  • SM-2 is still fine if you want intervals you can work out by hand, or if your app does not offer FSRS.
SM-2FSRS
What it modelsHow easy each card is, as an ease factor[1]Each card's memory: difficulty, stability and probability of recall[2][3]
InputsYour grade at each reviewYour grades and the time between reviews, across your whole review history
How the next interval is setPrevious interval × ease factorWhen the estimated probability of recall falls to your desired retention
PersonalisationPer card, through the ease factor; the same rules for everyoneParameters fitted to your own review history[3]
TransparencyA few lines of arithmetic you can check by handA model with 21 parameters (FSRS-6); intervals are hard to work out by hand
Where availableAnki (its legacy scheduler) and many other flashcard apps[4]Anki, AnkiMobile and AnkiWeb from 23.10, AnkiDroid from 2.17[4]
First released1987, by Piotr Woźniak in SuperMemo[1]2022, by Jarrett Ye, as a script for Anki[5]

What is the difference between FSRS and SM-2?

SM-2 sets each new interval by multiplying the previous one by an ease factor, while FSRS sets it from an estimate of how likely you are to recall the card. In SM-2, the first two intervals are 1 and 6 days; after that, each interval is the previous one times the card's ease factor, which starts at 2.5, drops when you answer with difficulty and never goes below 1.3.[1] If you forget a card, its intervals start again. The formula is detailed on the SM-2 page.

FSRS keeps three numbers per card: difficulty, stability and retrievability, the probability that you can recall the card now.[3] It comes from a memory model published by Jarrett Ye and colleagues at KDD 2022.[2] The model's parameters are fitted to your review history, and FSRS schedules the next review for when retrievability falls to your desired retention, 90% by default in Anki.[4] The FSRS page explains the model.

How are FSRS and SM-2 compared?

The main comparison is the open-source srs-benchmark, which checks how well each algorithm's estimated probability of recall matches what users actually remembered. It uses review logs from 10,000 Anki users and evaluates each algorithm on later reviews than the ones it learned from.[6] It is run by the FSRS developers, so it is not an independent test.

  • Log loss measures how far the estimated probabilities are from what happened at each review; lower is better.
  • RMSE (bins) groups similar reviews and measures the gap between the average estimate and the share actually recalled; lower is better.
  • AUC measures how well the algorithm tells apart the reviews you passed from the ones you failed; 0.5 is chance, higher is better.

SM-2 does not estimate probabilities itself, so the benchmark adds formulas based on assumptions about the forgetting curve to make it comparable.[7]

Which is more accurate, FSRS or SM-2?

FSRS is clearly more accurate: it beats both the original SM-2 and Anki's variant of it on all three metrics. The figures below come from the same 9,999 collections and about 350 million reviews, not counting same-day reviews.[6][7]

AlgorithmLog loss ↓RMSE (bins) ↓AUC ↑
FSRS-7 (latest version)0.3400.0630.717
FSRS-6 (used in Anki)0.3460.0650.703
Baseline: always guess the user’s average0.3950.1030.500
Anki’s SM-2 variant0.6160.1720.613
SM-20.7220.2030.603
FSRS and baseline rows: srs-benchmark README, read 8 November 2026. SM-2 rows: last README version that listed them, 24 July 2025; the current README no longer lists SM-2.

The gap holds for almost every user: FSRS-6 trained with more weight on recent reviews had a lower log loss than Anki's SM-2 variant for 99.5% of collections.[7] SM-2's log loss and RMSE are even worse than the baseline's, but SM-2 was never designed to output probabilities.[7]

Does FSRS mean fewer reviews?

Probably, but the evidence is weaker than for accuracy. The FSRS documentation says users “have to do 20–30% fewer reviews” than with SM-2 for the same retention, and adds that the figure is based on simulations.[3] We found no controlled study with real learners that measures it.

What FSRS does give you is a direct trade-off. In Anki, a higher desired retention means shorter intervals and more reviews; above 90% the workload rises quickly, and above 97% it can be overwhelming.[4] With SM-2 there is no retention setting; you change the workload indirectly, for example with Anki's interval modifier.[4]

When is SM-2 still a reasonable choice?

SM-2 is still a reasonable choice when predictability or simplicity matters more than accuracy. Cases where it may be fine:

  • You want to know the next interval in advance. SM-2 is a few lines of arithmetic you can check yourself.
  • You have few reviews logged. FSRS needs data to fit its parameters, and Anki warns that it may not perform well with fewer than a few hundred reviews.[4] Its default parameters still did better than SM-2 in the benchmark.[6][7]
  • Your app only offers SM-2. Many apps use SM-2 or a variant. A consistent review habit matters more than the algorithm behind it.

How do you switch from SM-2 to FSRS in Anki?

Turn on FSRS at the bottom of the deck options, click “Optimize” to fit the parameters to your history, and choose a desired retention.[4] Anki has used FSRS-6 since version 25.07, released in July 2025.[8] Leave “Reschedule cards on change” off when you first switch, as the manual advises, to avoid a sudden pile of due cards.[4] See the Anki algorithm for both schedulers in detail.

Sources

  1. Woźniak, P. A. (1990). Optimization of learning (Master's thesis, University of Technology in Poznań), section 3.2, published online as “SuperMemo 2: Algorithm”. https://super-memory.com/english/ol/sm2.htm
  2. Ye, J., Su, J., & Cao, Y. (2022). A stochastic shortest path algorithm for optimizing spaced repetition scheduling. In Proceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining (pp. 4381–4390). https://doi.org/10.1145/3534678.3539081
  3. Open Spaced Repetition. ABC of FSRS. awesome-fsrs wiki, GitHub. Retrieved 8 November 2026. https://github.com/open-spaced-repetition/awesome-fsrs/wiki/ABC-of-FSRS
  4. Anki manual. Deck options: FSRS. Retrieved 8 November 2026. https://docs.ankiweb.net/deck-options.html#fsrs
  5. Open Spaced Repetition. fsrs4anki v1.0.0-alpha (release, 20 September 2022). GitHub. https://github.com/open-spaced-repetition/fsrs4anki/releases/tag/v1.0.0-alpha
  6. Open Spaced Repetition. SRS Benchmark (README, results without same-day reviews, last updated 25 October 2026). GitHub. Retrieved 8 November 2026. https://github.com/open-spaced-repetition/srs-benchmark
  7. Open Spaced Repetition. SRS Benchmark (README, version of 24 July 2025, “Unweighted” results and “Superiority”). GitHub. https://github.com/open-spaced-repetition/srs-benchmark/blob/45f61b2182c27da83839383ec3b044fa5ae27b47/README.md
  8. Ankitects. Anki 25.07 release notes. GitHub, 4 July 2025. https://github.com/ankitects/anki/releases/tag/25.07