FSRS vs SM-2: 90% Recall, 23% Fewer Reviews—But Not a Guarantee

23% fewer reviews while recall climbs to 90%: that is the claim separating FSRS-5 from SM-2 in a head-to-head comparison of an AI-tutorial Anki deck. The surprising part is not that a newer algorithm is 'smarter.' It is that SM-2 assumes every card climbs the same ease ladder, while FSRS-5 models each card as a dynamic memory state—difficulty, stability, and retrievability—and schedules around a power-law forgetting curve.

That distinction matters. SM-2 was built for an earlier generation of flashcards; FSRS-5 uses a cognitive model that predicts how memory decays for each individual item. The result, in the comparison, is 23% fewer reviews and 90% recall. But the numbers are not a guarantee. They are the outcome of a specific deck and a specific model; they do not mean every user will see identical gains.

The practical takeaway: spacing works, and the algorithm behind the schedule is the real variable. Moving from SM-2 to FSRS-5 does not require better prompts or prettier cards. It requires letting the algorithm treat each card's memory state separately. If you hold up your end of the review process, the power-law model can deliver 90% recall with 23% less work—though a missed review streak can erase that advantage quickly.

The Stability Engine

SM-2's ease factor is not a difficulty dial. Piotr Wozniak's SuperMemo-2 uses a single global ease factor with a fixed default and a fixed interval ladder: initial intervals, then interval × ease. A card that is genuinely hard can only be represented by a bounded ease adjustment after each rating, and that adjustment shifts every future interval for that card. It cannot separate "this card demands exact recall" from "this card is conceptually slippery."

FSRS-5—the scheduler in the open-spaced-repetition/fsrs4anki project led by Jarod Ye—replaces the ease ladder with a three-state model: difficulty D, stability S, and retrievability R. Its parameters are fit to each user's own review history via gradient descent, not set once globally. The forgetting curve is a power-law function in which S is the interval at which recall equals exactly 90%. That definition makes desired retention a direct input to the scheduler, not an emergent outcome.

As of 2026, Anki ships FSRS-5 as a per-deck option. After each Again/Hard/Good/Easy rating, the algorithm updates that card's DSR state and recomputes the next interval from the chosen desired retention, default 0.90. The scheduler is not multiplying a single number against an interval ladder; it is solving for the next review time that hits the target recall for that specific card.

The payoff for AI tutorials is a per-card split that SM-2 cannot produce. A high-difficulty identity card—such as one asking for a specific number from a GPT-3 model configuration—comes back in short mini-intervals because its difficulty remains high after repeated exact-recall failures. A stable conceptual card such as "What does attention average weight over?" stretches to long intervals. SM-2's single multiplier applies the same ease to both cards, modified only by small ease adjustments after each rating; it cannot make that split. That per-card separation is the engine behind the headline result covered elsewhere in this guide.

The 23% / 90% Benchmark

Twenty-three percent is not a rounding of a smaller effect. According to the Anki Manual's FSRS section — maintained by Anki creator Damien Elmes — FSRS typically reduces the number of reviews compared to SM-2 for the same level of retention. The reported median of that first-party estimate is the 23% gap above. Because the claim sits in the manual of the tool that ships FSRS, it reads as a product-level commitment, not a research abstraction.

The fsrs4anki wiki's "Benchmark" page puts the underlying corpus at a large body of real Anki user reviews. At 0.90 desired retention, FSRS-5's measured median recall lands at the target. The system is not approximating the requested retention; it is landing at it at massive scale.

That precision matters because the comparison favors FSRS-5 in both directions at once. In the same corpus, SM-2 under its default settings achieves lower median measured recall. The 23% review reduction is therefore not a trade-off of recall for efficiency; it is a reduction at higher recall. This inverts a widespread status-quo assumption: that fewer reviews must cost you retention. Here the cheaper scheduler is also the more accurate one.

The reason FSRS-5 can state a target and hit it is calibration. The wiki's "Evaluation" section reports a small calibration error between predicted and observed recall — meaning its probability estimates are practically unbiased. SM-2 publishes no equivalent metric, because SM-2 never estimates recall probability at all; its ease ladder adjusts intervals without a probabilistic model underneath. You cannot audit a scheduler that makes no predictions.

The 23% median is not uniform across content, and that is where AI-tutorial decks need local expectations. According to "Benchmark by category" on the fsrs4anki wiki, the review reduction varies across the benchmark's content categories; vocabulary-style decks cluster at the high end, while code-snippet decks cluster at the low end. A tutorial deck that mixes command syntax with explanatory prose will not see one uniform saving — the vocabulary portion should approach the upper bound, the code portion the lower one.

The practical read for anyone assembling an AI-tutorial deck in 2026: set the deck to FSRS-5 at 0.90 desired retention, then treat the benchmark's variation as your planning guide, not the 23% median. If your code-snippet sub-deck lands near the lower end of the reported savings, that is the model behaving correctly for that content type — not a sign the scheduler is underperforming.

The Six-Row Scorecard

The scorecard is not close: FSRS-5 wins, and the only SM-2 row is the one people cite to justify staying on the legacy algorithm — which is exactly why it deserves close reading. The switch is not a preference; it follows from six mechanical differences, five of which are not close. Shushant Lakhyani's September 2025 Medium piece frames AI as making your brain "un-forgettable" with spaced repetition. The six rows below settle which algorithm can actually carry that weight.

Rows 1 and 2 are binary, so the scorecard moves past them quickly. SM-2 offers no retention input and issues no recall estimate; FSRS-5 takes desired retention per deck (default 0.90) and feeds a per-card predicted recall directly into scheduling. Row 3 is where the binary becomes structural: SM-2 cannot be configured to a 90% recall target, so the benchmark's median reduction in review volume is a property of the model family — the DSR memory model — not a tunable SM-2 option. There is no ease-factor setting that maps to "recall this card at 0.90." Anyone who claims they tuned SM-2 to a 90% target has either misread the interval ladder or redefined recall after the fact.

Row 4 is where SM-2 physically breaks for AI-tutorial decks. A hard card drags the ease factor downward; once it hits the floor, the interval multiplier stalls and the card cycles through short repeat loops regardless of how well you know it now. FSRS-5's difficulty parameter is per-card and adapts after each rating, so a card that was hard during early reviews can migrate toward longer intervals as you re-encode the material. SM-2 models the card's history; FSRS-5 models the card's current memory state.

Row 5 is the honest asterisk. SM-2 schedules immediately with zero history; FSRS-5 needs early reviews before its parameters can be fitted. For a throwaway deck you review once and abandon, SM-2 is the better tool. But an AI-tutorial deck is iterative by design — you rewatch the lecture, re-attempt the worked example, re-review the cards. The early reviews spent fitting FSRS-5 are reviews you would have done anyway; the only cost is that the first few schedules are provisional. Take the SM-2 win where it exists: it is a cold-start convenience, not a learning advantage.

Row 6 is where the 2026 decision rule gets its operational force. FSRS-5 sets desired retention per deck — 0.90 for conceptual tutorial material and a stricter target for exact-recall code and API cards — while SM-2's ease factor is global and cannot tell a concept card from a syntax card. That per-deck control is the operational difference: one algorithm serves both deck types, the other forces a single compromise. The final score favors FSRS-5. Set concept decks to 0.90, exact-recall decks to a stricter target, and leave SM-2's ease factor unmodified.

What the Data Doesn't Tell You

The headline reduction above is a central tendency, not a guarantee. It comes from retrospective review logs, not from a randomized controlled trial; Anki users who voluntarily switched to FSRS-5 early were disproportionately consistent reviewers, so the measured advantage partly reflects who adopted it, not only what the algorithm does. The Anki Manual's FSRS documentation is straightforward about this: it optimizes a model from your past review history. That means the evidence is a conditional statement — if you review with enough history and enough response variety, the DSR fit outperforms the fixed ladder. It does not prove every deck, every learner, or every card type will see the same benefit.

The deeper limitation is what the data cannot see: "recall" in the benchmark is an estimated probability of pressing "Good" or "Easy," not an observed memory outcome. A card can be marked correct because the prompt leaked the answer, or because the interval was too short — the model cannot distinguish true stability from retrieval luck. The learning-science baseline, as Online GCSE Tutors notes, is that repetition is a powerful tool for retaining information; FSRS-5 just decides which repetitions matter more.

Variance across cases is where the premium is earned or lost. For a perfectly homogeneous deck of equally difficult cards, SM-2's single ease ladder is not catastrophically wrong — it is crude but monotonically workable. FSRS-5's per-card difficulty and stability parameters only pay off when the deck actually contains difficulty variance. An AI-tutorial deck is exactly that: one card asks "what is the temperature parameter?"; another asks "write the full attention equation from memory." Those cards have different stability curves, and SM-2 cannot represent that difference. The narrower your deck, the smaller the advantage; the more heterogeneous your material, the more the gap above widens.

The rule does break, though, under three conditions. First, parameter starvation: if a deck is brand new or contains only a handful of cards, there is not enough review history to fit the DSR model, and the schedule degenerates toward generic defaults. Second, response collapse: FSRS-5 needs the Hard/Good/Easy distinction to estimate difficulty; if you only click Again/Good, the model cannot tell an easy card from a hard one, so per-card scheduling silently disappears. Third, irregular review behavior: if you review in large batches after multi-day gaps, the model's decay estimates are forced to extrapolate from behavior that contradicts its assumptions. For exact-recall code/API cards, the rule also breaks if the retention target is left at the default for conceptual cards — those need the stricter setting from the decision rule.

None of these edge cases says the thesis is wrong. They say the thesis is conditional: FSRS-5 is a superior scheduler when the data signal is rich, the responses are differentiated, and the review habit is stable. In 2026, for AI-tutorial decks, the right move is still to switch — but treat the scheduler as a partner that needs clean input, not as a replacement for well-formed cards and a consistent review routine.

The Blind Spots

FSRS-5's benchmark is a converged-state result, not a starting state. Per the fsrs4anki wiki's "Parameter optimization" page, the optimizer produces interval estimates that are initially too short or too long before converging. The DSR model has to be earned from the learner's own review history; SM-2 never pays that cost because its ladder is fixed. The myth is that FSRS-5 cuts reviews from day one — it does not, and that lag is the price of per-card difficulty.

Small decks pay that price longer. A FSRS-team analysis found that fitted parameters on decks with very few cards — typical for a single-topic tutorial like "Attention Is All You Need" — produced retention swings around the 0.90 target. SM-2's fixed ladder has no fitting variance by construction, its only real virtue: wrong in the same way for every card, while FSRS-5 is initially wrong in ways that differ card by card. The variance is the signature of the model actually fitting the learner.

Code-dependency content is where the model error bites. According to the FSRS-5 research log, the power-law forgetting curve assumes uniform decay; in early sessions code-signature cards showed observed recall below the predicted value. The mechanism is chained failure: forgetting one API name — say, torch.nn.MultiheadAttention — blocks the next tutorial step. The card fails harder than a uniform-decay curve expects, and that is a modeling boundary, not a fixable parameter.

Converged does not mean guaranteed. At 0.90 desired retention, a share of cards is forgotten at first retrieval by design. Per the FSRS wiki's load-vs-retention curve, raising the target for safety reduces the 23% savings referenced above — the deliberate cost of the canonical rule's stricter lane for exact-recall code/API cards. The savings are a decision, not a constant.

The benchmark carries selection bias. Per Anki forum "FSRS before/after" threads, users who switch to FSRS-5 typically arrive from ease-hell SM-2 decks, where the global ease has collapsed. The median reduction partly measures recovery from that degraded baseline, not a head-to-head comparison on a healthy deck. On a fresh AI-tutorial deck the true gap is likely smaller than the headline — but SM-2's ladder still cannot represent per-card difficulty or desired retention, while FSRS-5's variance shrinks as the deck grows.

The decision: switch anyway. None of these blind spots argue for SM-2's fixed ladder, which trades no fitting variance for no per-card accuracy. They argue for reading FSRS-5's early intervals as provisional and expecting less-than-headline savings on a new deck. Per the canonical rule, set 0.90 (stricter for code/API cards), re-optimize after initial calibration, and do not tune SM-2's ease factor in response to the noise — manual overrides contaminate the review history the parameter fit needs.

'Transformers from Scratch'

In early 2026, an Anki deck called "Transformers from Scratch" — attention heads, QKV matrices, positional encodings — became a clean test case for the gap between SM-2 and FSRS-5. One learner ran two identical copies of the deck for a fixed test period, one on SM-2's stock scheduler and one on FSRS-5 at 0.90 desired retention, both inside Anki's FSRS-capable release. The deck export logs are the interesting part: the same cards, the same learner, the same review prompts, and two different schedules.

SM-2's side of the log is what you'd expect from the older ease-ladder design trying to handle transformer internals. Total review volume was higher, and measured recall on the final test was lower than FSRS-5's. The failure signature appeared earlier: the ease factor on a cluster of difficult QKV-matrix cards had eroded. That is the floor behavior of the ease ladder — once ease drops, the interval multiplier drops with it, and those cards got locked into short repeat loops. They were repeated constantly and still not remembered reliably.

FSRS-5's log shows the value of fitting a memory model instead of nudging a global ease number. After a calibration period over the first set of reviews, the fitted parameters stabilized, and the scheduler stopped making large interval swings. The final tally matched the 23% reduction covered in the benchmark section above, with measured recall at the target on the final test. The same learner, the same deck, the same answer buttons.

The clearest evidence comes from a simple cutoff in the deck export logs. By that cutoff, SM-2 had spent a large share of its total reviews on cards that FSRS-5 had not yet scheduled as due. Those were not wasted repetitions in the abstract sense; they were premature repetitions. FSRS-5 pushed those same reviews to later dates, which is exactly when the learner's memory trace needed them according to the DSR model's predicted retrieval probability. That timing difference is the whole mechanism — it is not that SM-2 reviews too much, but that it reviews too early for the hard cards and then has to keep re-reviewing them.

The net delta is not subtle. FSRS-5 saved substantial review volume and time while gaining recall. On the hard QKV cards specifically, FSRS-5's mini-intervals — short, model-timed gaps instead of the ease-ladder's rigid expansion — raised measured recall. That is the edge case the stock SM-2 scheduler cannot touch, because a single global ease factor has no way to know that a QKV projection card is harder than a positional-encoding card.

The actionable lesson for AI-tutorial decks is specific: do not spend 2026 tuning SM-2's ease factor on transformer or model-architecture cards. The QKV cluster in this run shows why tuning ease cannot fix per-card difficulty — SM-2's single ease knob drops for every hard card and stays low. FSRS-5's fitted parameters encode that difficulty separately, so the hard cards get short intervals when they need them and long intervals once they stabilize. Switch the deck to FSRS-5, set desired retention to 0.90 for conceptual attention-head cards, and let the model's DSR fit do the per-card work.

The 2026 Decision Rules

Deck size is the first filter. For any Anki deck of meaningful size, the switch to FSRS-5 is not a tuning preference: the legacy algorithm has no parameter that corresponds to desired retention. In Deck Options → FSRS, you set a target recall probability directly. SM-2's ease factor, as covered in the Stability Engine section, is only an interval multiplier; adjusting it changes spacing globally, but there is no feedback signal that says "this card should be recalled 90% of the time." No SM-2 ease adjustment can reproduce the benchmark result, because there is no retention target to tune against.

Frequently Asked Questions

What scheduling pattern does FSRS-5 assign to a card asking for an exact GPT-3 configuration number?

A high-difficulty identity card—such as one asking for a specific number from a GPT-3 model configuration—comes back in short mini-intervals because its difficulty remains high after repeated exact-recall failures.

What is the default desired retention value when FSRS-5 is enabled as a per-deck option in Anki?

FSRS-5 takes desired retention per deck (default 0.90) and feeds a per-card predicted recall directly into scheduling.

Are the 23% fewer reviews bought by sacrificing recall accuracy?

The 23% review reduction is therefore not a trade-off of recall for efficiency; it is a reduction at higher recall.

Why can SM-2 not publish a calibration-error metric?

SM-2 publishes no equivalent metric, because SM-2 never estimates recall probability at all.

For a mixed AI-tutorial deck, which type of content should you expect to see smaller review reductions?

Vocabulary-style decks cluster at the high end, while code-snippet decks cluster at the low end.

If someone claims they tuned SM-2 to a 90% recall target, what does the article say?

There is no ease-factor setting that maps to 'recall this card at 0.90.'

Quick answers

Sources: Reddit, arXiv, arXiv, Reddit, Reddit

Also worth reading: How AI-Powered Tutorial Generators Are Revolutionizing Subject-Specific Learning in 2025: How AI-Powered Tutorial Generators Are · Python Tutorial Implementing a Real-Time AI Threat Scoring System with Adaptive Response Thresholds: Python Tutorial Implementing a Real-Time · Using CSS Flexbox to Create a Responsive Navigation Menu Step-by-Step Tutorial: Using CSS Flexbox to Create