Why FSRS Is Better Than SM-2

Most flashcard users never look at what algorithm schedules their reviews. They tap "Good," see the next card, and move on. But the scheduling algorithm matters more than the app's interface, its price, or how many decks it ships with. It determines which cards you see, when you see them, and whether you waste twenty minutes a day on reviews you don't need.

SM-2 was groundbreaking in 1987. It gave every card its own schedule — something no software had done before. But it has structural flaws that compound over months of use, and FSRS, published 35 years later, solves them. In this article, I'll walk through how both algorithms actually work, where SM-2 breaks down, and what the published benchmarks show.

What SM-2 Actually Does

Piotr Wozniak created SM-2 for SuperMemo in 1987. It was the first algorithm to personalize review intervals for individual cards instead of using a fixed schedule for an entire deck. At the time, this was a genuine breakthrough — most students were using paper flashcards with no system at all, or Leitner boxes with fixed intervals.

Here's how SM-2 works. Every card gets a single number called an ease factor (EF), starting at 2.5. After each review, the algorithm adjusts this number based on your response quality (a 0-5 scale, though most modern implementations use 1-4). The next interval is calculated by multiplying the previous interval by the ease factor:

I(n) = I(n-1) x EF

So a card with EF = 2.5 and a current interval of 10 days gets scheduled for 25 days. If you rate it "Good" again, it goes to 62 days, then 156. The growth is exponential, which is correct — memory strength does grow exponentially with successful retrievals.

The problem is what happens when you get a card wrong.

Each incorrect or difficult answer reduces the ease factor. Get a card wrong, and EF drops — typically by 0.2 or more. The minimum is clamped at 1.3, meaning intervals can never grow faster than 30% per review. But here's the structural flaw: there's almost no mechanism to raise EF back up. Getting a card right only adds a small amount (0.1 at most), and only if you rate it "Easy" — which most people rarely do.

The result is a one-way ratchet. A card you struggle with twice early on drops to EF = 1.3 and stays there. You review it every 2-3 days, then every 4 days, then every 5 — even after you've genuinely learned it. This is "ease hell": cards trapped at minimum intervals, consuming review time that should go to cards you actually need to practice.

The ease factor acts as both a measure of card difficulty and a scheduling multiplier — two different concepts fused into one number. That fusion is the root cause of SM-2's problems, and it's exactly what FSRS decouples.

What FSRS Does Differently

FSRS (Free Spaced Repetition Scheduler) was created by Jarrett Ye and published in 2022. The original paper appeared at the ACM SIGKDD 2022 workshop, with a more comprehensive version published in IEEE Transactions on Knowledge and Data Engineering in 2024. Unlike SM-2, FSRS was developed in the open, with public benchmarks and open-source code from day one.

The fundamental difference: FSRS uses three parameters instead of one.

• Stability (S): How long, in days, until your recall probability drops to 90%. A stability of 30 means you'd have a 90% chance of remembering the card after 30 days. This is purely about memory strength — how deeply the card is encoded.
• Difficulty (D): How inherently hard the card is for you, on a 1-10 scale. This reflects the card itself — an abstract grammar rule vs. a concrete noun, for example. Difficulty changes slowly and captures the card's intrinsic challenge.
• Retrievability (R): Your current probability of recalling the card right now. This decays over time following a forgetting curve calculated from stability. When R drops to around 0.9 (90%), FSRS schedules a review.

The key insight is that stability and difficulty are independent. A hard card (high D) that you've successfully reviewed 50 times can still have very high stability — FSRS can represent this. SM-2 cannot. In SM-2, a card that was ever difficult has a permanently low ease factor, so its intervals stay short regardless of how many times you've gotten it right since.

FSRS also models a per-card forgetting curve rather than applying a single multiplier. The scheduling target is the point where your retrievability drops to approximately 90% — challenging enough to strengthen the memory, but not so late that you've already forgotten.

And because FSRS learns from your personal review history, the model adapts to you over time, not population averages.

The Numbers — How Much Better Is FSRS?

Claims need evidence. The FSRS papers include extensive benchmarks against SM-2 and other algorithms, tested on review histories from over 10,000 Anki users.

The key metric is log-loss — a measure of how well the algorithm predicts whether you'll remember or forget a card. Lower log-loss means more accurate predictions, which means better scheduling.

What does this mean in practice? More accurate predictions translate directly to fewer wasted reviews. When the algorithm correctly predicts you'll remember a card, it doesn't schedule an unnecessary review. When it correctly predicts you're about to forget, it catches the card before you lose it. The difference between SM-2's and FSRS's accuracy adds up to measurably fewer daily reviews for the same retention rate — or higher retention for the same daily review load.

The benchmarks are fully reproducible. The FSRS optimizer is open-source, and anyone can run the benchmark on their own Anki history. This level of transparency is rare in the SRS space — most apps treat their algorithm as proprietary, even when it's just SM-2 with minor tweaks.

Why the Default Matters

Anki added FSRS as an option in late 2023 — a significant move, since Anki is the largest SRS platform in the world. But there's a catch: it's buried in Advanced Settings, disabled by default, and requires users to explicitly opt in. Most Anki users never change advanced settings. They study with SM-2 without knowing a better option exists.

This creates a paradox: the best scheduling algorithm is technically available to millions of users, but the vast majority never benefit from it because the default hasn't changed.

Defaults shape behavior at scale. Research in behavioral economics consistently shows that opt-out systems have adoption rates of 80-90%, while opt-in systems sit at 20-30% — even when the option is clearly beneficial. The same principle applies to SRS algorithms. An algorithm that could help isn't the same as one that does help.

Words on Repeat ships FSRS as the default for every user — free and paid. No configuration screen, no advanced settings toggle, no YouTube tutorial explaining where to find it. You sign up, pick a deck, and start studying with state-of-the-art scheduling from your first review. The algorithm is invisible, which is exactly how it should be.

What This Means for Your Study Sessions

Switching from SM-2 to FSRS changes your daily experience in ways you feel before you understand why:

• Fewer pointless reviews. Cards you've mastered stop appearing unnecessarily. SM-2 users often report reviewing 30-50% more cards per day than FSRS users at the same retention level.
• Hard cards eventually graduate. Instead of being stuck at minimum intervals forever, cards that were once difficult can reach long intervals after you demonstrate consistent recall. The ease-hell death spiral doesn't exist in FSRS.
• Predictions match reality. When the algorithm says a card is due, it's actually due — not just arbitrarily scheduled by a formula with no feedback mechanism. This reduces the friction that makes people quit their study habit.
• It just works. The best algorithm is one you never have to think about. FSRS adapts to your review patterns automatically, without you tweaking parameters or worrying about optimal settings.

For the deeper science behind why spaced repetition works at all — the forgetting curve, encoding variability, and 140 years of memory research — see The Science of Spaced Repetition.

Frequently Asked Questions

References

1. Wozniak, P.A. (1990). Optimization of Learning. supermemo.com — Original description of the SM-2 algorithm.
2. Ye, J. (2022). "A Stochastic Shortest Path Algorithm for Optimizing Spaced Repetition Scheduling." ACM SIGKDD Workshop on Deep Learning for Search and Recommendation. doi:10.1145/3534678.3539081
3. Su, J., Ye, J., Nie, L., Cao, Y. & Chen, Y. (2024). "Optimizing Spaced Repetition Schedule by Capturing the Dynamics of Memory." IEEE Transactions on Knowledge and Data Engineering. doi:10.1109/TKDE.2023.3251721
4. Open Spaced Repetition. FSRS4Anki — Free Spaced Repetition Scheduler. github.com/open-spaced-repetition/fsrs4anki — Open-source optimizer and benchmark data.
5. Anki (2023). Changes in Anki 23.10 — FSRS integration announcement. changes.ankiweb.net