Every spaced-repetition app is really two things: a set of cards, and an algorithm that decides when to show each one. For more than thirty years that algorithm was almost always SM-2, the scheduler Piotr Woźniak published for SuperMemo in 1987 and that Anki ran, in modified form, for most of its life. In late 2023 Anki added a new one, FSRS, and made it the recommended default. If you have seen "FSRS vs SM-2" and wondered whether the upgrade is real or just a new acronym, this is the honest explanation.
The short answer
SM-2 is a fixed heuristic: a simple rule of thumb, tuned by hand in the 1980s, that grows each card's interval by an "ease" number it nudges up or down based on how you rate yourself. FSRS is a memory model fit to data: it predicts the probability that you will recall a given card at a given moment, and schedules the review for when that probability drops to a retention target you choose.
Because FSRS is fit to a public benchmark of more than 700 million real reviews, it predicts recall more accurately than SM-2. In practice that accuracy buys you the same retention with roughly 20 to 30 percent fewer reviews, or higher retention for the same number of reviews.
One line: SM-2 guesses your forgetting with a 1987 formula; FSRS measures it with a model trained on hundreds of millions of reviews.
How SM-2 works, and where it strains
SM-2 tracks three numbers per card: an interval, a repetition count, and an "ease factor" that starts around 2.5. Each time you pass a card, the next interval is the current one multiplied by that ease factor, so intervals grow geometrically: a few days, then a couple of weeks, then a month, and so on. When you rate a review, the ease factor moves up a little if it felt easy and down if it felt hard, with a floor around 1.3. Fail a card and it drops back into learning.
For a heuristic invented before most learners were born, this is remarkably good, and it is why Anki worked so well for so long. But it has known structural limits:
- One number does too much. A single ease factor has to capture how hard a card is for you, how durable the memory currently is, and how much the next interval should stretch. Those are different things, and squeezing them into one value loses information.
- It cannot model memory directly. SM-2 has no explicit notion of how likely you are to remember a card right now. It just multiplies intervals. So it cannot aim at a target like "keep me at 90 percent recall."
- It was tuned, not learned. The constants come from one person's experiments decades ago, not from your reviews or anyone else's at scale. It cannot adapt to the fact that your cards, or you, forget differently.
- Ease hell. Repeatedly rating cards "Hard" drives the ease factor down and never lets it recover well, so some cards get stuck reviewing far too often. Anki users have a name for this because it happens a lot.
How FSRS works
FSRS is built on the DSR model of memory, which describes any card with three quantities:
- Difficulty: how hard it is to make this particular card stick for you.
- Stability: how long the memory will last, defined as the time for your recall probability to fall from 100 percent to 90 percent.
- Retrievability: the probability you can recall the card right now, which decays as time passes since the last review.
Instead of one hand-tuned ease number, FSRS uses a set of parameters, around twenty in recent versions, that it fits to review history so the model's predicted recall matches what actually happens. With that model, scheduling becomes a precise question: review each card at the moment its retrievability drops to the retention you asked for. Set a desired retention of 90 percent and FSRS spaces each card to be due right as its predicted recall hits 90 percent, no sooner and no later. You can fit those parameters to your own history, so the schedule adapts to how you specifically forget.
The practical effect is that FSRS stops wasting reviews on cards you would still remember and stops letting cards you are about to forget slip too far.
The evidence: a benchmark, not a vibe
The reason this is not just a nicer story is that FSRS is measured. Its developers maintain an open benchmark that pits FSRS against SM-2 and other schedulers on a public dataset of more than 700 million reviews from tens of thousands of real users, scoring each algorithm on how well its predicted recall matches the actual outcome. FSRS predicts recall more accurately than SM-2 across that dataset.
Better prediction is what turns into fewer reviews. Because FSRS knows more precisely when you are about to forget, it does not pad every interval with the safety margin a blunter algorithm needs, which is where the commonly reported 20 to 30 percent reduction in reviews for the same retention comes from. That is a real, reproducible result, not marketing.
FSRS vs SM-2, side by side
| SM-2 | FSRS | |
|---|---|---|
| Origin | SuperMemo, 1987, hand-tuned | Open-source, fit to data, 2022 onward |
| Model of memory | None; multiplies intervals by an ease factor | DSR: difficulty, stability, retrievability |
| Parameters | A few fixed constants | Around twenty, fit to review data |
| Targets a retention level | No | Yes; you pick the desired retention |
| Adapts to your forgetting | Barely | Yes, from your own history |
| Reviews for the same retention | Baseline | Roughly 20 to 30 percent fewer |
| Known failure mode | Ease hell | Far less prone to it |
| Where to use it | Older Anki defaults | Anki since v23.10, and Anti-Agent |
Does this mean you should switch?
If you use Anki, switching is trivial and worth it: FSRS has shipped in Anki since version 23.10 and is the recommended scheduler. Turn it on in the deck options, let it optimize on your existing review history, and your future intervals get better with no change to your cards. The gains are largest for big, mature decks with lots of review history to learn from, and more modest for small or brand-new decks where there is little data to fit.
If you use Anti-Agent, you are already on FSRS by default, so there is nothing to switch. We did not invent our own scheduler, because the right move on the algorithm question is to use the one that wins the benchmark. If you want the head-to-head on the tools rather than the algorithms, we wrote that up separately in Anti-Agent vs Anki.
Where the algorithm stops mattering
Here is the honest limit of this whole debate. FSRS is better than SM-2, clearly. But the best scheduler in the world only decides when to show you a card. It cannot decide whether you actually recalled it. That part is still up to how you review.
In both Anki and classic SM-2, you flip the card and rate yourself, which means the algorithm is fed your own judgment of whether you knew it. And people are generous: seeing the answer and thinking "yes, that's roughly what I had" feels like recall but usually is not. It is the fluency illusion, the same reason notes and reviews can feel productive and still evaporate. Feed a perfect scheduler optimistic self-ratings and it will faithfully schedule you to forget.
That is the gap we care about most. A great algorithm plus honest grading beats a great algorithm alone, which is why our cards ask you to type the answer and have it graded against the source rather than self-rated. FSRS decides the timing; grading decides the truth.
Frequently asked questions
Is FSRS actually better than SM-2, or just newer? Better, and measurably so. On a public benchmark of more than 700 million reviews, FSRS predicts recall more accurately than SM-2, which translates to roughly 20 to 30 percent fewer reviews for the same retention.
What does SM-2 stand for? It is the second SuperMemo algorithm, published by Piotr Woźniak in 1987. Anki used a modified version of it for most of its history before adopting FSRS.
Does Anki use FSRS or SM-2 now? Both are available, but FSRS has been built in since Anki v23.10 and is the recommended scheduler. Older collections that have not switched still run the SM-2-based default.
Do I lose my cards or history if I switch to FSRS in Anki? No. FSRS uses your existing review history to optimize its parameters. Your cards stay exactly as they are; only the future scheduling changes.
Is the algorithm the most important thing for remembering? It matters, but less than people think. FSRS controls when you review. Whether the review actually tests recall, rather than letting you recognize an answer and rate yourself generously, matters at least as much.
The bottom line
FSRS beats SM-2 because it replaced a clever 1987 rule of thumb with a memory model fit to hundreds of millions of real reviews, so it schedules each card for the moment you are about to forget it instead of guessing. If you are on Anki, enable it. If you are on Anti-Agent, you already have it.
Just remember the scheduler is only half the job. It picks the moment; you still have to actually recall. Try the version where your answer gets graded, not flipped, and see how that changes what comes back.
Anti-Agent vs Anki: an honest comparison
A direct, brand-aware comparison of Anti-Agent and Anki. They run the same FSRS scheduler, so the real choice is manual cards you flip and self-rate versus generated cards an AI grades. Here is where each one wins, and when to use both.
NotebookLM has no spaced repetition. This alternative is built around it.
NotebookLM is great at chatting with your sources, but nothing comes back and nothing tests you. Here is a NotebookLM alternative with flashcards and spaced repetition, where your sources turn into graded recall that resurfaces on a schedule.