FSRS vs SM-2: Why Your Scheduling Algorithm Matters

If you have used Anki, you have been scheduled by SM-2. If you have used a newer system with "optimised scheduling," you may have encountered FSRS. Both algorithms serve the same purpose — deciding when to show you each flashcard next — but they do it in fundamentally different ways. The difference matters more than most learners realise, because your scheduling algorithm determines how much time you spend reviewing, how well you actually retain material, and how the system responds when your memory behaves unexpectedly. For context on all the spaced-repetition options available, browse the spaced repetition guide collection.

SM-2: The Classic Heuristic

SM-2 was developed by Piotr Wozniak in the 1980s for SuperMemo and has been the default scheduling algorithm in Anki since its creation. Its design reflects the constraints of its era: it needed to be simple enough to run on early personal computers and explainable enough for users to understand manually.

SM-2 works roughly as follows. Each card carries an ease factor — a multiplier that begins at a default value (typically 2.5) and adjusts based on your ratings. After the initial two learning steps (which use fixed short intervals), the next interval is computed by multiplying the previous interval by the ease factor. If you rate a card Hard, the ease factor decreases, shortening future intervals. If you rate it Easy, it increases slightly, extending them. A card you repeatedly struggle with gradually accumulates a low ease factor and gets reviewed more frequently; a card you consistently nail gets pushed further into the future.

This design works well enough to be genuinely useful — millions of learners have built substantial knowledge using SM-2. But it has structural limitations that become apparent when you study at scale or analyse the underlying assumptions.

Key Limitations of SM-2

• Ease factor drift. Once an ease factor decreases — through Hard or Again ratings — it tends to stay low. Cards can get "trapped" in short intervals even after the learner has clearly internalised the material, leading to over-review of stable memories.
• No explicit forgetting model. SM-2 does not predict when you will forget; it applies a heuristic multiplier to the previous interval. The resulting schedule may or may not align with actual memory decay for a given card and learner.
• Uniform decay assumption. The algorithm assumes all cards decay at similar rates, adjusted only by the ease factor. It has no concept of different stability levels arising from the history and pattern of past reviews.
• Cold-start rigidity. Every new card starts with the same default ease factor regardless of how similar content has been handled in the past. The system does not learn your patterns across cards.

FSRS: A Memory Model, Not a Heuristic

FSRS — the Free Spaced Repetition Scheduler — takes a different starting point. Instead of defining rules for how intervals should change, it builds a mathematical model of memory itself and derives intervals from that model. The algorithm is open source and peer-reviewed; see the FSRS open-source repository for the technical specification.

As described in more detail in How FSRS Works, the model tracks three quantities for each card: difficulty (intrinsic hardness for this learner), stability (how slowly the memory decays), and retrievability (probability of successful recall at this moment). The next review is scheduled for the point in time when retrievability is predicted to reach a target threshold — commonly 90 %.

Because the model makes explicit predictions about recall probability, it can set intervals that are long when memory is genuinely stable and short when it is fragile — decisions that SM-2 can only approximate through its ease-factor mechanism.

Head-to-Head Comparison

Property	SM-2	FSRS
Core mechanism	Ease-factor heuristic applied to the previous interval	Memory model predicting retrievability over time
Interval derivation	Previous interval × ease factor	Time until retrievability drops to target threshold
Retention targeting	Implicit — no explicit retention probability	Explicit — configurable target (e.g. 90 %)
Per-learner adaptation	Per-card ease factor adjustments only	Model weights can be fitted to individual review history
Ease factor drift	Yes — low ease factors tend to persist	Not applicable — stability replaces ease factor
Rating signal used	Adjusts ease factor and interval multiplier	Updates stability and difficulty; re-derives interval
Transparency	Simple rules, easy to explain manually	More complex model, but explicit retention predictions
Review efficiency	Good for most use cases	Generally fewer reviews for equivalent retention

Why the Algorithm Affects Review Efficiency

The review-efficiency difference between the two algorithms comes down to how well each one tracks the actual state of your memory. SM-2 uses a proxy — the ease factor and its accumulation over past reviews — that correlates with memory stability but does not measure it directly. The result is a schedule that is reasonable on average but imprecise for individual cards.

FSRS, by tracking stability explicitly, can distinguish between two situations that SM-2 treats similarly: a card with a low ease factor that you have recently consolidated through a run of Good reviews, and a card with a low ease factor that remains genuinely fragile. SM-2 will give both roughly the same short interval; FSRS will give the consolidated card a longer interval because its stability has grown.

Over a large deck studied for months, this precision compounds. Stable memories are not reviewed unnecessarily; fragile memories are reviewed before they drop below the retention threshold. The net effect is a review queue that is smaller and better targeted, which means less time spent reviewing while maintaining the same — or higher — actual retention.

Trade-offs and Practical Considerations

SM-2 is not without advantages. It is extremely transparent: you can read the algorithm in a few lines and understand exactly what it will do next. It has decades of real-world validation. And for learners with small decks or short study timelines, the efficiency difference between SM-2 and FSRS may be negligible in practice.

FSRS introduces more complexity. The model has parameters that need sensible defaults — and ideally, calibration to your data. It requires a consistent rating discipline, because inaccurate ratings mislead a model more than they mislead a heuristic. For a learner who rates cards carelessly or uses a binary pass/fail approach, FSRS's advantage shrinks significantly.

The other important consideration is what "retention" means in each system. SM-2 does not give you a retention target to set; the resulting retention depends on the default ease factor and how you rate cards. FSRS lets you choose your target explicitly — 70 %, 90 %, 95 % — and adjusts intervals to maintain it. Higher targets mean shorter intervals and more reviews; lower targets mean longer intervals and higher forgetting. Having explicit control over this trade-off is useful when you have strong opinions about how reliable your recall needs to be, such as for a professional certification versus casual language learning.

Choosing Between Them

If you are starting fresh and your study system supports FSRS, there is little reason to use SM-2. The model-based approach is more accurate, more adaptable, and produces better outcomes at scale. The learning curve for the user is minimal — you still rate cards the same way; the algorithm just uses those ratings more intelligently.

If you have an existing SM-2 deck with years of review history, migrating to FSRS may or may not be straightforward depending on your tool. Some systems can estimate FSRS stability from SM-2 history; others start fresh. The migration path is worth checking before committing.

For anyone serious about active recall as a long-term practice, the algorithm matters increasingly as the deck grows. A few hundred cards reviewed casually may not reveal the difference; tens of thousands of cards studied over years makes every efficiency gain compound meaningfully.

The best scheduling algorithm is the one you will actually use consistently. But given equal commitment, FSRS produces more accurate schedules, more explicit retention control, and better efficiency at scale.

Scheduling Algorithms and the Broader Learning Loop

It is worth remembering that scheduling is one component of a learning system, not the whole of it. The quality of the flashcards you are reviewing, your engagement during initial reading, the depth of understanding you bring to each recall attempt — these factors interact with the scheduling algorithm to determine actual learning outcomes. A mediocre card scheduled perfectly is still a mediocre card.

The value of FSRS is that it removes the scheduling bottleneck. Once the algorithm is accurately modelling your memory and setting intervals appropriately, the remaining variables are in your control: how you read, how you generate cards, how honestly you rate, and how consistently you show up for reviews.

LearnBook uses FSRS throughout — so when you are reviewing material across multiple sources at once, every card in the cross-book queue carries its own independently calibrated interval rather than being forced onto a shared schedule. To see FSRS in practice within the full study workflow, read about how flashcards and quizzes are generated and scheduled.