How Flashside schedules: FSRS 6, the weights, the steps

FSRS is an open spaced repetition algorithm. Version 6 has 21 weights, a forgetting curve whose shape is learned rather than fixed, and formulas for answers given on the same day. Flashside implements it in its own core library, without a scheduler package; the phone app and the web run the same functions.

Moving from FSRS 4.5

No deck is re-planned. A deck switches at its first answer on a build that has FSRS 6, and the switch syncs to your other devices. The due dates it already has stay where they are; a card's stored stability and difficulty are read by FSRS 6 at its next answer, and only answers given after the switch use the new formulas. The memory chart replays each answer on the formulas that graded it: before the switch on FSRS 4.5, after it on 6.

These are the published FSRS 6 defaults, pre-trained on a large set of open review logs and not fitted to any Flashside learner yet. w20 is the decay, the shape of the forgetting curve, which FSRS 4.5 fixed at -0.5 across 17 weights. w17 to w19 move a card's stability on an answer given the same day, which is what the learning steps are.

The target retention is the probability of recall each review is scheduled at: by default a card comes back on the day your chance of still knowing it falls to 90%. You can set it from 0.7 to 0.97. A higher value shortens the intervals and brings more reviews; a lower one lengthens them.

After each answer you tap one of four grades. Each moves the card's stability (in days) and its difficulty. The last column is what each button shows under it for a card you meet for the first time.

A new card is seen again after 1 minute and 10 minutes before it enters the day-based schedule. A card you forget in a review waits 10 minutes before it is tried again.

The second step is your own: Flashside moves it along 5 · 10 · 20 · 30 minutes from your answers on it. Once there are 40 answers on it in the last 60 days, a pass rate of 95% or more lengthens it by one notch and 80% or less shortens it by one, at most once in 14 days. The step for a forgotten card follows it.

With FSRS 6, these same-day answers feed the model through w17 to w19; FSRS 4.5 left them out.

The Meet card. When a deck introduces new cards, a card you have never seen is shown first as a Meet card: you guess, see the answer and tap Got it. That puts it on its first learning step without a grade, so the first answer FSRS reads is your first real recall.

Exam mode. From 21 days before an exam date, every card you have met comes back once a day until you have recalled it on 3 separate days. Those reviews come ahead of the interval FSRS would give, and the date ends the mode.

The plan's new-card rate. The study plan, not FSRS, decides how many new cards a deck releases a day: from your minutes a day and exam date, or 20 by default. FSRS schedules a card once it is introduced.

“Learn N more”. When today's new cards are done, a deck offers another day's worth. The extra cards take the learning steps and FSRS like any other new card, and tomorrow's queue stays as it is.

Personal weights. The 21 weights are the same for every learner. A fit to your own reviews is built and checked on synthetic logs; it ships only once at least 3 real learners have 1,000 reviews over 90 days, counted monthly from 15 January 2027. A fit on less would be a guess.

FSRS is published by the open-spaced-repetition project together with the open SRS Benchmark, which compares schedulers on public review logs. Your four grades, the time between your answers, the target retention and the weights above are everything the scheduler reads.

The memory chart's “you would recall today” is the mean FSRS recall probability over the deck's started cards.

Sources: the FSRS algorithm · the SRS Benchmark