FSRS vs. SM-2: What the Research Actually Shows About Anki's New Algorithm
Published: September 8, 2026 | By: Roman Vasylyshyn, founder of Lambdio | Reading time: 9 min
If you've used Anki in the last two years, you've probably seen a banner suggesting you switch your deck to "FSRS." If you've never used Anki, you may not know that spaced repetition software has quietly gone through its biggest algorithm change in over 30 years.
For three decades, almost every spaced repetition app — Anki, SuperMemo's early versions, Mnemosyne, most flashcard apps — ran on some variant of a single algorithm: SM-2, published by Piotr Wozniak in 1987. In 2023, Anki added support for a fundamentally different algorithm called FSRS (Free Spaced Repetition Scheduler), and by 2024 it had become the recommended default.
This post explains what actually changed, why it matters for how long you spend studying, and what the evidence says — without the marketing spin that usually surrounds "AI-powered" learning claims.
Table of Contents
- The Problem SM-2 Was Trying to Solve
- Why SM-2 Started Showing Its Age
- What FSRS Does Differently
- The Three-Component Memory Model
- Side-by-Side: SM-2 vs. FSRS
- What the Benchmarks Actually Show
- Does This Matter for You?
- How Lambdio Uses FSRS
The Problem SM-2 Was Trying to Solve
Every spaced repetition algorithm is trying to answer one question: given that you just reviewed a fact, when should you see it again?
Too soon, and you waste time reviewing something you'd have remembered anyway. Too late, and you've already forgotten it, so the review does little good. The entire discipline of spaced repetition is about finding the interval that sits right at the edge of forgetting — where recall is effortful but still usually successful.
SM-2, designed by Piotr Wozniak for the original SuperMemo software, solved this with a simple heuristic: track a single "ease factor" per card. Every time you rate a card (Again, Hard, Good, Easy), the ease factor nudges up or down, and the next interval is calculated as a multiple of the previous one. It's elegant, computationally trivial, and it's why spaced repetition became practical on 1980s hardware.
It's also why it's remained the default in most flashcard software for over 35 years — it works well enough, and it's easy to implement.
Why SM-2 Started Showing Its Age
SM-2's core limitation is that it treats forgetting as roughly the same process for every card and every learner. In reality:
- Some material is inherently harder to retain than other material, independent of how well you rated it last time.
- Memory stability doesn't grow linearly. Each successful recall makes a memory more durable, but with diminishing returns — the tenth successful review doesn't help as much as the third.
- Learners differ. Two students with identical review histories on the same card don't necessarily forget it at the same rate.
SM-2 has no explicit way to represent any of this. It has one adjustable number per card (the ease factor) and fixed rules for how that number moves. It can't be optimized — there's no way to feed it data and have it learn better rules, because it isn't a model of memory, it's a fixed procedure.
This is the gap FSRS was built to close.
What FSRS Does Differently
FSRS — Free Spaced Repetition Scheduler — is an open-source algorithm developed by Jarrett Ye and the open-spaced-repetition community. Instead of a hand-tuned heuristic, FSRS is a genuine mathematical model of memory, with parameters that are fit to real review data rather than guessed by a designer decades ago.
The practical difference: SM-2 has 1 tunable value (the ease factor). FSRS has 19 trainable weights that together model how difficulty, stability, and time interact to determine forgetting.
Those weights were optimized against the FSRS-Anki-20k dataset — roughly 740 million individual review events collected from about 20,000 real Anki users. That's the key shift: FSRS isn't just a better guess, it's a model fit to an enormous amount of real human forgetting behavior.
The Three-Component Memory Model
FSRS is built on a model with three variables per card, derived from Piotr Wozniak's own later work on the three-component model of memory (yes — the same researcher who created SM-2 in 1987 went on to describe the more nuanced model that FSRS implements decades later):
| Component | What it represents |
|---|---|
| Difficulty (D) | How inherently hard this specific piece of material is to retain, on a 1–10 scale |
| Stability (S) | How resistant the memory currently is to decay — grows after successful recall, resets after failure |
| Retrievability (R) | The estimated probability you could recall this right now, at this exact moment — decays continuously with time |
Every review updates all three. A successful "Easy" rating increases stability more than a "Good" rating, but the size of that increase depends on the current difficulty and stability — mirroring the diminishing-returns pattern seen in real memory research. A failed "Again" rating drops stability, with the size of the drop depending on how stable the memory had become.
The next review is then scheduled by calculating exactly when retrievability is predicted to fall to a target threshold — by default, 90%. That's a meaningfully different design philosophy from SM-2's fixed interval multipliers: FSRS is explicitly solving for a probability of recall, not applying a formula.
Side-by-Side: SM-2 vs. FSRS
| Aspect | SM-2 (1987) | FSRS |
|---|---|---|
| Model type | Fixed heuristic | Trained mathematical model |
| Tunable parameters | 1 (ease factor) | 19 weights |
| Difficulty tracking | Indirect, folded into ease factor | Explicit variable (D) |
| Stability growth | Linear multiplier | Non-linear, with diminishing returns |
| Forgetting modeled explicitly | No | Yes — via retrievability (R) |
| Target recall probability | Fixed (~85–90% implicitly) | Configurable (default 90%) |
| Can be optimized on data | No | Yes — weights can be retrained per dataset or per user |
| Basis | Designer's intuition | 740M+ real review events |
What the Benchmarks Actually Show
It's worth being precise here rather than hand-wavy, because "AI-powered" claims in edtech are cheap and a lot of them don't hold up.
The open-spaced-repetition team's published benchmarks — run against real user review logs, not synthetic data — consistently show FSRS achieving one of two outcomes compared to SM-2: higher same-day recall for the same number of reviews, or fewer total reviews required to hit the same retention target. In other words, for a fixed amount of study time, FSRS schedules reviews closer to the actual point of forgetting, which is the entire goal of spaced repetition in the first place.
The algorithm's design and benchmarking methodology have been published in peer-reviewed venues, including ACM KDD 2022 and IEEE TKDE 2023.
It's also worth naming the limits honestly. FSRS's advantage is largest when there's enough review history to estimate difficulty and stability well — a brand-new card with one or two reviews doesn't give the model much to work with yet, so early-stage scheduling is closer to SM-2's territory of "reasonable default." And because FSRS is a statistical model rather than a fixed rule, it can occasionally schedule an interval that feels wrong for an individual card in a way a human wouldn't — most implementations let you manually adjust when that happens.
Does This Matter for You?
If you're spending 20 minutes a day on flashcards, the difference between SM-2 and FSRS probably isn't going to be dramatic in any single week. Where it compounds is over months and years of studying — which is, not coincidentally, the actual timescale spaced repetition is designed for.
Two practical takeaways:
- If you use Anki, switching your deck's scheduler to FSRS (Anki 23.10+) is close to a free upgrade — it uses your own review history to fit the weights, so it gets more accurate the longer you've been using the deck.
- If you're evaluating any spaced repetition tool — including Lambdio — "we use spaced repetition" is no longer a meaningful differentiator by itself. The more useful question is which algorithm, and whether it's actually been benchmarked against real forgetting data or just described as "AI-powered" without specifics.
How Lambdio Uses FSRS
Lambdio's scheduling engine is a custom implementation blending FSRS v5 and v6, built on the same open-source reference specification described above. Every learning and review session — Standard, Socratic, Chat Review, Feynman, or Quiz mode — ends with an Again/Hard/Good/Easy rating that feeds directly into the FSRS engine, which recalculates difficulty, stability, and the next review date for that chapter.
A few implementation specifics, for anyone curious:
- Desired retention is set to 90% by default — the same target used in the FSRS benchmarks above.
- Card states are tracked per chapter (New, Learning, Review, Relearning), so a chapter you've mastered and later forgotten is scheduled differently from one you're seeing for the first time.
- Quiz scores are automatically mapped to FSRS ratings (85%+ → Easy, 70–84% → Good, 50–69% → Hard, below 50% → Again), so the algorithm updates even during multiple-choice review sessions, not just conversational ones.
The scheduling math is the same regardless of which learning mode you use — what changes across modes is how you engage with the material before that rating gets recorded. We covered the pedagogical research behind those modes (testing effect, generation effect, desirable difficulties) in the platform overview.
If you want to see the scheduling in action against real course material, courses are free to start at lambdio.com.
Sources: FSRS4Anki Wiki, FSRS-Anki-20k dataset, Three Component Model of Memory.