If you've spent any time with flashcard apps, you've probably come across terms like "SM-2" or "FSRS" without anyone really explaining what they mean. You nodded, moved on, and hoped it didn't matter. It does — a little. These are the scheduling algorithms that decide when you see each card again, and choosing the right one is the difference between reviewing cards at the perfect moment and reviewing them so early you've barely had time to forget them (or so late that you absolutely have).
Where did these algorithms come from?
SM-2 (SuperMemo 2) is the older of the two. It was created by Polish researcher Piotr Woźniak in 1987 for his SuperMemo learning program and updated in 1989. It spread quickly and was adopted by apps like Mnemosyne and Anki (in a modified form) starting in 2006. Decades later, many apps still use SM-2 or close variants — a testament to how well a simple idea can age.
FSRS (Free Spaced Repetition Scheduler) is much newer. It was released as open source by developer Junyao Ye ("Jarrett Ye") in September 2022, then integrated as a native scheduling option in Anki in late 2023 — which drove widespread adoption almost overnight. In algorithm terms, it's basically a newborn. A very impressive newborn.
How are they different?
SM-2 applies a simple, consistent logic. Each time you answer a card correctly, the interval before you see it again gets longer. Each time you struggle, it shortens. The key variable is called the "Ease Factor," which adjusts based on your performance over time. Think of it like rowing a boat toward shore: a good stroke brings you closer, a missed stroke lets the current pull you back out to sea. It works reliably — but it treats every rower the same, regardless of how tired their arms are.
FSRS starts from a different premise. Rather than a single ease factor, every card is tracked with three variables:
The three FSRS variables
Stability
How long you can retain the card before forgetting it. Higher stability means a longer interval before the next review.
Difficulty
Harder cards get shorter intervals so they come back sooner. No card gets to hide.
Retrievability
The probability that you'll remember the card right now. When R drops below a threshold, the card gets rescheduled — before it's gone for good.
The algorithm also learns from your personal review history and improves its scheduling over time. Using the same boat metaphor: your vessel is now equipped with sensors measuring current strength, wind direction, and your own fatigue — each stroke timed for maximum effect. Less effort, better results. The dream.
Which one is actually better?
FSRS generally outperforms SM-2 on the metric that matters most: achieving the same retention with fewer total reviews. Once it has enough data to calibrate to you — typically after a few hundred reviews — it produces meaningfully smarter schedules.
That said, SM-2 is far from obsolete. Its logic is simple, reliable, and starts working immediately without needing to learn your patterns first. FSRS, by contrast, needs data before it really shines. In the early days of a new deck, the difference may be imperceptible — or FSRS might even feel slightly worse, which is a mildly humbling experience for an algorithm that's supposed to be the future.
The practical answer: if you're using a modern app that supports FSRS and you're willing to stick with it past the initial calibration period, it's the stronger long-term choice. Patience rewarded.
Why Noos uses FSRS
When we built Noos, we spent a long time evaluating which algorithm to implement. SM-2 was the obvious safe choice — battle-tested, well-documented, and familiar to most learners. We chose FSRS instead, for one reason: language learning is exactly the context where its adaptability matters most.
Vocabulary cards are not created equal. The Russian word for "and" (и) and the word for "nevertheless" (тем не менее) have completely different difficulty profiles — and they decay at different rates depending on the person. FSRS tracks that individually. SM-2 applies a universal ease factor and hopes for the best. For a language with 6 cases and 3 genders, hoping for the best isn't really a strategy.
If you're learning a language seriously, you want an algorithm that models your memory, not an average learner's. That's what FSRS does — and it's why it's at the core of Noos.