FSRS Explained: How an Algorithm Predicts When You'll Forget

Yesterday, a flashcard felt obvious. Today, you hesitate. Next week, it might be gone—or still come back immediately. A useful scheduler has to work with that uncertainty rather than assume every correct answer deserves the same delay.

TL;DR: FSRS estimates how a memory changes, predicts the probability of recalling a card after a delay, and chooses a review interval for a target recall probability. Higher targets generally mean earlier reviews. The prediction is a model, not a promise about your next answer.

This guide follows that decision from memory state to review date. If you want the shorter introduction first, our beginner’s guide to FSRS explains the basic vocabulary.

Why review timing needs a memory model

Fixed schedules cannot distinguish different memories

A schedule such as “review after one day, three days, then a week” is easy to follow. It also treats a familiar word and a confusing definition alike. Even when both answers are correct today, their chances of surviving a week can differ substantially.

The problem is not that fixed schedules are useless. They provide structure and space retrieval attempts across time. But a growing deck contains memories at different stages, and repeated reviews of secure cards compete with the cards you genuinely struggle to retrieve.

Adaptive scheduling uses your answers as evidence

FSRS stands for Free Spaced Repetition Scheduler. It uses a card’s review history, including the elapsed time and ratings, to update an estimated memory state. It then turns that estimate into the next interval.

The important distinction is between updating a card’s state and optimizing the model’s parameters. A review can update the card immediately using the current parameters. Fitting parameters to a learner’s accumulated history is a separate process supported by some applications; using FSRS does not automatically mean an app performs personalized optimization.

Nor are all older schedulers fixed ladders. SM-2 also adapts intervals using answer quality. FSRS makes a different modeling choice: it explicitly estimates memory stability and recall probability. That is the useful comparison, rather than “old algorithms never adapt.”

How FSRS models your memory

Difficulty: how hard the card is for you

Difficulty represents the estimated challenge of a particular card for a particular learner. A short definition can be difficult if it resembles three other definitions. A longer answer may be easy because you already understand the idea behind it.

In FSRS, difficulty helps determine how stability changes after reviews. It is not simply a timer that says “hard card equals tomorrow.” Two difficult cards can still have very different intervals if one memory is much more stable.

Stability: the time scale of forgetting

Stability has a precise meaning in the model: the number of days until predicted recall falls to 90% after a review. A card with ten-day stability therefore has a modeled 90% chance of recall after ten days, under the assumed forgetting curve.

It does not mean the memory vanishes on day eleven, and it is not the same thing as the chosen review interval. With the same stability, a scheduler can choose an earlier interval for a higher desired retention. Successful retrieval often increases estimated stability; failure supplies evidence for a different update.

Retrievability: the estimated chance of recall now

Retrievability is the probability of recalling the answer at a given time. It depends on how long it has been since the review and on the estimated stability. Difficulty and stability describe aspects of the memory; retrievability answers the immediate scheduling question.

Think of two vocabulary cards: one has a stability of two days, the other twenty days. After ten days, they need not have comparable recall probabilities, even if you last answered both correctly. The same elapsed time represents a much larger challenge for the first memory.

These definitions come from the FSRS algorithm documentation. They describe a predictive model of review outcomes, not a direct measurement of what your brain contains.

From forgetting curves to review dates

A worked example: one memory, three targets

Hold stability at ten days and imagine no intervening reviews. The model starts with high recall probability and predicts a gradual decline. A review threshold is the point where this curve reaches your desired retention.

For the illustrative curve below, an 80% target gives an interval of about 24 days. At 90%, the interval is ten days. At 95%, it is about 4.6 days. The memory estimate has stayed the same; only the acceptable predicted risk at review has changed.

That also explains a surprisingly long interval: it can reflect a strong estimated memory, a lower target, or both. A date alone does not tell you whether the app thinks a card is difficult. You need the memory estimate and scheduling settings to interpret it.

Explore the retention threshold

Change the target and watch the threshold move along the same curve. The horizontal line represents your requested recall probability; its intersection with the curve gives the approximate interval.

Explore FSRS retention targetsIllustrative demo

Target recall probability

80%90%95%

Approximate review threshold

10

days

90%

· Target recall probability

Fixed ten-day stability; FSRS-4.5 curve. No repeated reviews or personal forecast.

This demonstration uses the documented FSRS-4.5 forgetting curve, with fixed ten-day stability, to isolate one idea. It does not simulate repeated reviews, fit personal parameters, or reproduce a particular app’s schedule. Newer FSRS versions can use a different curve shape, and real applications may round intervals, add small scheduling variation, or apply learning steps and interval limits.

Notice that selecting a higher target does not make the curve stronger. It moves the review earlier. Improving the underlying memory requires learning and retrieval; the target only changes when you plan to check it again.

What changes after a review

When the next review happens, the scheduler receives another observation: the elapsed time and your rating. It updates difficulty and stability, then chooses the next interval from the updated state. The next curve can therefore differ from the one you saw before.

A correct answer after a substantial delay can provide different evidence from an immediate repeat. A failure also matters: it tells the model that recall did not succeed under those conditions. This is why a repeated-review schedule cannot be represented honestly by extending our single fixed-stability curve forever.

You can think of the process as a cycle:

1. Try to recall before revealing the answer.
2. Rate the outcome honestly.
3. Update the estimated memory state.
4. Choose the next interval for the desired retention.
5. Repeat with new evidence.

Choosing a retention target

Why 90% is a common starting point

Desired retention is the probability you want the model to target when a card is due for review. Anki uses 90% as its default and describes it as a balance between retention and workload. It is a starting point, not a scientifically universal optimum for every learner or task. (Anki deck options)

A target of 90% does not mean you permanently sacrifice one tenth of your knowledge. Cards have different ages at any moment, and a failed review becomes an opportunity to relearn. Nor does the target guarantee that exactly nine of your next ten answers will be correct. Actual outcomes fluctuate, and an imperfectly fitted model may be poorly calibrated for your material.

Look at trends across enough reviews to judge a setting. A difficult afternoon, a few newly learned cards, or a changed deck is a weak basis for declaring the model accurate or inaccurate.

Higher targets bring a workload trade-off

Raising the target shortens intervals for the same memory state. In the illustration, moving from 90% to 95% more than halves the interval. That is a comparison of one theoretical interval, not a claim that your total review workload will exactly double.

Long-term workload also depends on new cards, failures, changes in stability, learning steps, and your available study time. The Anki manual warns that workload rises quickly above 90%, especially at very high targets. Choosing 99% because it sounds almost perfect can produce a schedule you cannot sustain.

Use the consequence of forgetting to guide the decision. Long-term vocabulary practice and a near-term exam may justify different priorities. If the workload becomes unmanageable, consider fewer new cards and better card design alongside the retention setting. Increasing a target cannot create extra time in your day.

FSRS vs SM-2

Different approaches to the next interval

The original SM-2 algorithm adjusts an ease factor from answer quality and uses it to grow intervals after successful repetitions. Failed responses restart the repetition sequence. It adapts, but it does not explicitly select intervals by solving for a requested recall probability.

FSRS models difficulty and stability, derives retrievability over elapsed time, and chooses an interval around a retention target. Its parameters can be optimized against review histories where an application supports that workflow. Both systems still depend on feedback from reviews.

Be careful with app comparisons: an application described as “SM-2 based” may use substantial modifications. Likewise, two apps using FSRS can differ in version, parameters, learning steps, interval limits, and review behavior. Algorithm names alone do not specify the full user experience.

Strengths and limitations in practice

FSRS offers an explicit way to reason about predicted recall and the trade-off between retention and timing. That makes it easier to explain why a review is due and how a target affects intervals. Parameter fitting can also adapt a model to recorded review patterns.

Its limitations remain practical. Sparse or misleading history, inconsistent ratings, ambiguous cards, and material learned outside the app can all make prediction harder. A model sees logged evidence, not every conversation, lecture, or practice problem that strengthened your memory.

There is no need to promise a fixed percentage reduction in reviews to explain the value. Better prediction can help allocate reviews more usefully; whether that improves your workload and recall depends on your data, settings, and habits.

Making reviews fit real life

Busy days and overdue cards

A due date is a model-based recommendation, not a biological expiry date. If you miss it, predicted recall generally declines further, but the card does not instantly become forgotten. Review honestly when you return so the scheduler gets evidence about what happened after the longer delay.

Reserve realistic review time in your calendar and reduce new-card intake when a backlog grows. A calendar slot organizes your time; the scheduler organizes individual cards. Those are complementary tasks, and using FSRS does not by itself imply a calendar integration.

SpaceRep uses FSRS to schedule flashcard reviews and applies deadline-aware scheduling for exam decks. Treat those dates as support for a study routine while keeping time for understanding and application.

Exam deadlines change the planning problem

An exam creates a second constraint: you need useful recall by a particular date, even if a long-term scheduler would place a review later. Apps can handle that with higher retention targets or interval limits, but the behavior is implementation-specific.

Do not spend every available minute chasing a perfect flashcard statistic. Written explanations, mixed problems, and timed practice test skills that a compact recall prompt cannot cover. Our guide to cramming versus spaced repetition before an exam shows how to balance review with the assessment’s actual demands.

Honest ratings and clear cards improve the evidence

If you could not recall the answer, report a failure. In FSRS, “Hard” describes successful but difficult recall; it should not become a polite substitute for “Again.” Otherwise, the history suggests success where none occurred.

Also check the card itself. Is there one clear question? Can you tell whether your answer was correct? Does it test understanding or merely a familiar-looking sentence? Our guide to effective flashcard design helps you fix prompts that make ratings unreliable.

Good scheduling starts with meaningful observations. A well-designed card, an actual retrieval attempt, and an honest rating give the model something useful to work with—and make the review worthwhile even when the prediction is imperfect.

FAQ

Does 90% retention mean forgetting 10% of everything?

No. It is a target recall probability at the scheduled review time, not a permanent quota of lost knowledge. Cards reviewed more recently can have higher predicted recall, and failed cards can be relearned.

Does FSRS learn my personal parameters after every review?

Not necessarily. Each review updates the card’s memory state using the current parameters. Optimizing parameters from your history is a separate feature and depends on the application.

Can FSRS guarantee that I remember something on exam day?

No. It estimates probabilities from review evidence. Deadline-aware scheduling may help you plan, but recall also depends on the quality of learning, the assessment task, and conditions on the day.