FSRS vs SM-2: Which Spaced Repetition Algorithm Is Better?

If you use flashcards to learn anything — Japanese vocabulary, kanji, medical terms, whatever — there's a quiet piece of software deciding when each card comes back for review. That algorithm decides how much you remember, how many reviews you do per day, and whether your habit feels sustainable or soul-crushing. For roughly forty years that algorithm was SM-2, the scheduling method Piotr Woźniak wrote for SuperMemo in 1987. Since Anki 23.10, a newer algorithm called FSRS has been available as a built-in option — you have to turn it on, but it's one toggle away — and the flashcard world has been arguing about it ever since.

Here's what actually changed, how each algorithm works, and which one you should use.

What spaced repetition is, in one paragraph

Every memory you form decays on a curve: right after you learn something you remember it perfectly, then it slips — fast at first, then more slowly. This is the forgetting curve first described by Hermann Ebbinghaus in 1885. A spaced repetition system (SRS) interrupts that decay by asking you to recall the item right before you'd forget it. Each successful recall makes the memory more stable, so the next review can wait longer. The algorithm's entire job is to estimate that moment — when, exactly, will I almost forget this card? — and schedule the review for just before it. Review too early and you waste time; review too late and you've already forgotten it.

SM-2: the 40-year-old workhorse

SM-2 was created in 1987 by Piotr Woźniak for SuperMemo (original algorithm description), and it's the algorithm that powers Anki and has powered most flashcard software for decades. It's genuinely elegant in its simplicity.

Each card carries two numbers: an interval (how many days until the next review) and an ease factor (a multiplier that represents how easy the card is for you). When a card comes up, you grade yourself on a four-button scale:

• Again — you forgot it. Interval resets to zero, the card comes back soon, and the ease factor drops.
• Hard — you got it, but it was a struggle. Small interval increase, ease factor drops slightly.
• Good — normal recall. Interval multiplied by the ease factor.
• Easy — instant, effortless recall. Bigger interval increase, ease factor rises.

The strengths of SM-2 are real: it's simple, it's predictable, it's been battle-tested by tens of millions of learners over four decades, and the math is simple enough to fit on a napkin. It works. The weaknesses, though, became obvious over time:

• It doesn't actually model memory. SM-2 has no notion of "how strong is this memory right now" — it just multiplies intervals. The ease factor is a crude proxy for memory strength, and it doesn't correspond to any real cognitive quantity.
• Ease hell. Because the ease factor only drops and recovers slowly, a few early "Hard" or "Again" grades on a card can pin its ease factor so low that it comes back every day forever, flooding your review queue. The card isn't hard; it's just stuck.
• One curve for everything. SM-2 assumes the same forgetting-curve shape for every card and every person. A tricky kanji and an easy word get the same mathematical treatment, just with different multipliers.

FSRS: the memory-science successor

FSRS — the Free Spaced Repetition Scheduler, developed by Jarrett Ye and collaborators starting in 2022 — takes a fundamentally different approach. Instead of multiplying intervals, it models memory directly using three variables borrowed from cognitive science:

• Stability — how durable the memory is. A stable memory survives a long time before decaying; an unstable one fades in days.
• Difficulty — how hard this specific item is for you, independent of stability. Some cards are just stubborn.
• Retrievability — the probability you can recall the card right now, at this moment, given how long it's been since you last saw it.

FSRS uses these to predict the exact point at which your recall probability will hit a target retention rate — 90% by default, meaning the algorithm tries to bring each card back when you have a 90% chance of still remembering it — and schedules the review for that moment. You can dial the target up (say, 95% if you're studying for a high-stakes exam and want more safety margin) or down (85% if you want fewer reviews and can tolerate more failures).

The practical payoff: FSRS typically delivers the same or higher retention at fewer reviews than SM-2. It also avoids the ease-hell trap entirely, because there's no ease factor to spiral — difficulty is a real modeled quantity that adjusts smoothly with your actual performance.

The key difference: personalization

Here's the single biggest reason FSRS outperforms SM-2: it adapts to you.

SM-2 uses the same curve shape for every card and every learner. A pre-med student cramming anatomy and a retiree learning French get the identical mathematical treatment. FSRS, by contrast, fits its parameters to your personal review history. After you've accumulated enough reviews — the FSRS docs suggest the optimizer needs at least a few hundred, and stabilizes with more — FSRS has enough data to optimize its memory model to your specific forgetting patterns: how fast you forget new cards, how much a single review strengthens a memory for you personally, which kinds of cards give you trouble. From that point on, every schedule it produces is tuned to your brain, not a generic average.

This is why FSRS pulls ahead over time. The first thousand reviews look similar to SM-2, but after optimization the schedules genuinely fit how you actually forget and remember.

In one line: SM-2 schedules every learner the same way; FSRS learns your forgetting curve and schedules to it.

Which retains more?

In benchmark testing, FSRS wins — not by a huge margin, but by a consistent and meaningful one. The FSRS project's benchmark suite (which evaluates the algorithm on large sets of real review logs) consistently reports FSRS achieving higher retention at lower review load than SM-2. The exact size of the win depends heavily on the dataset and the target-retention setting, so we won't quote a single percentage here — the takeaway is the direction (same-or-better recall for same-or-fewer reviews), not a precise number. (These are the developers' own results on their own data; independent academic replication is still thin, which is why this section stays qualitative.)

That's real, but it's not magical. On a 50-card daily deck, a modest review reduction is a handful of cards saved per day — worth having, not life-changing. The honest framing: FSRS is a genuine improvement, but it's an incremental one on top of a method that already worked well.

Which should you choose?

Practical guidance:

• Use FSRS if you're a typical learner starting fresh. It's the algorithm most Anki power users have moved to for good reason — same or better retention, fewer reviews, no ease hell, and it personalizes to you over time. (It's a one-toggle setting in Anki 23.10+; SM-2 remains the out-of-the-box default for now.) There's almost no downside.
• Stick with SM-2 if you're importing a long-standing Anki deck and want it to behave exactly as it always has. Some learners have years of tuned ease factors and don't want their schedules to shift when they flip algorithms. SM-2 is also the right call if you specifically want the classic, predictable behavior and don't care about optimization.
• In yomion, you can pick per deck. New material? FSRS. A deck you've been grinding for two years and don't want to disrupt? Leave it on SM-2. There's no globally correct answer — match the tool to the deck.

If you're unsure, default to FSRS. You can always switch later; your review history carries over either way.

The honest caveat

Here's the truth that gets buried in algorithm debates: the scheduler matters far less than consistency.

A learner who shows up every single day and grinds SM-2 cards will outperform a learner who uses FSRS sporadically, every single time, by a wide margin. The algorithm optimizes the last 10–15% of efficiency. Whether you show up at all determines the other 85%. Flashcard apps are full of people who've spent more time reading about scheduling algorithms than actually doing their reviews — algorithm-shopping is a seductive form of procrastination, because it feels productive while producing nothing.

Pick one. Turn it on. Stop reading about it. Go do your reviews.

The bottom line

SM-2 is a forty-year-old workhorse that has carried millions of learners to fluency and still works fine. FSRS is its modern successor — more accurate, personalized to your memory, and the algorithm the Anki community has overwhelmingly moved to. For most people starting today, FSRS is the right choice (you have to flip it on once, but that's it). For everyone, the choice that actually matters is showing up tomorrow.

Spaced repetition works because memory works that way. Pick the algorithm, build the habit, and let the curve do its job.

Sources

• SuperMemo — Application of a SuperMemo algorithm (SM-2 description) — Piotr Woźniak's original SM-2 algorithm writeup.
• Wikipedia — SuperMemo — history of SM-2 (1987, Woźniak) and its use across flashcard software.
• Wikipedia — Forgetting curve — Ebbinghaus's 1885 work underlying all spaced repetition.
• Wikipedia — Spaced repetition — overview of SRS techniques and history.
• AnkiWeb FAQ — What spaced repetition algorithm? — Anki's own statement on SM-2 vs FSRS (since 23.10, both are available).
• open-spaced-repetition/fsrs4anki (GitHub) — the FSRS project, maintained by Jarrett Ye and collaborators.
• FSRS — The Algorithm (wiki) — the Stability/Difficulty/Retrievability memory model FSRS uses.
• FSRS — Target retention (wiki) — how the 90% default (and 75–97% range) works.
• FSRS — Benchmarks (wiki) — the FSRS team's review-load comparisons against SM-2.
• Anki manual — Deck options / FSRS — how to enable FSRS (opt-in; SM-2 remains the out-of-the-box default) and the optimizer's minimum review count.
• AnkiWeb FAQ — What spaced repetition algorithm? — Anki's statement that SM-2 is the default and FSRS is available as an alternative since 23.10.
• Anki GitHub issue #3616 — Make FSRS the default? — the maintainer discussion about whether/when to flip FSRS to default (still open as of this writing; FSRS has not yet become the default).