FSRS in plain English

What the scheduler predicts

FSRS stands for Free Spaced Repetition Scheduler. It predicts the probability that you will recall a card at a given moment, and then chooses the earliest date at which that probability falls to your target retention. At the default 90%, it aims for a one-in-ten chance of forgetting.

That prediction is updated after every review. The scheduler is not a fixed ladder of intervals; it is a model of this card, for this person, right now. Two cards learned on the same day can diverge within a week if one keeps coming back forgotten.

Stability and difficulty

Two numbers do most of the work. Stability estimates how long the memory lasts before recall drops to the target, so it grows with each successful review. Difficulty estimates how hard the material is for you, so it grows when a card keeps lapsing and shrinks when it comes back easily.

A high-stability, low-difficulty card can wait a year. A low-stability, high-difficulty card returns in days. Both are in the same deck, and that is the point: a single global interval would be wrong for at least one of them.

New cards start optimistic and are corrected quickly. The first few reviews matter disproportionately because they set the initial estimate, which is why a calm first week beats a hundred cards added on a Sunday night.

None of this requires a decision from you. The scheduler recalibrates quietly, and the only input it needs is an honest grade.

Why Miyu asks only Again or Good

More buttons do not produce more information. They produce more optimism: a card you half-remember gets “Hard” when it should get “Again”, and the interval stretches anyway. Miyu shows two choices and the interval each one schedules, so the honest answer is also the easy one to give.

If you genuinely cannot recall the card, Again is not a failure. It is the signal that keeps the model accurate. A scheduler fed flattering grades will confidently schedule a card you have already forgotten.

What the first month feels like

New cards return quickly at first while the learning steps run, then the gaps open up: days, then a week or two, then a month. A small queue can look suspiciously easy in week three. That is the model working, not the deck running out.

Expect some churn in the first fortnight as the scheduler learns which cards are difficult for you. Adding a large batch of new cards at once delays that calibration; adding a few at a time lets the estimates settle.

If a card keeps coming back in days despite your best answers, the material may be too dense for one card. Splitting it is usually more effective than lowering retention for the whole deck.

Tuning retention

Retention, maximum interval, and learning steps are preset values you can edit. Raising retention above 90% buys fewer lapses at the cost of more reviews; lowering it trades accuracy for time. Change one value at a time and give it a few weeks.

Interval fuzz spreads reviews that would otherwise land on the same day. Leaving it on keeps the daily queue even, which matters more for consistency than any single tuning choice.

The defaults are a reasonable place to start precisely because they are boring. Tune after a month of real reviews, not before your first session, and change the preset that belongs to the affected deck rather than everything at once.

A useful experiment is to change nothing for two weeks and read the interval previews at the end of them. If Again appears far more often than Good, retention or card design needs attention; if Good dominates and the queue feels trivial, the preset is being generous.