The probability of recalling a card t days after the last review, given its stability.

FSRS-7 mixes two power laws whose relative weight depends on the stability itself, so — unlike every earlier version — the retrievability at t == s is not a fixed 0.9 but drifts with s.

t must be non-negative and s strictly positive.

d/dt of retrievability. Never positive for in-bounds parameters.

The interval, in days, after which a card of the given stability will have decayed to exactly the desired retention.

The FSRS-7 forgetting curve has no closed-form inverse, so this is a root-find: Newton's method in log t (which is what makes the problem well-conditioned — see the upstream fsrs_v7_interval_penalty module), safeguarded by a bracketing bisection so that it cannot diverge.

The result is always within [minimumIntervalDays, maximumIntervalDays]; a desired retention that is unreachable within that window saturates at the nearer end.

The difficulty a card is born with, given the rating of its first review.

Difficulty after a review: a rating-driven step, damped so that it slows down as difficulty approaches its maximum, then reverted 1% of the way towards the difficulty an Easy first review would have produced.

The stability a card is born with, given the rating of its first review.

One half of the stability update.

Called twice per review — once with longTermWeights and once with shortTermWeights — and the two results are blended by nextStability.

On a lapse the new stability is the post-lapse stability, which can never exceed the old one. On a success it is the old stability scaled by a factor that grows with how overdue the card was, shrinks as the card gets easier and more stable, and is scaled again by the hard penalty or the easy bonus.

How much of a long-term review this is: 0 for a review that happens at the same instant as the previous one, tending to 1 as the gap grows.

This is the piece that replaces FSRS-6's hard same-day/not-same-day split. The elapsed time is expected to be non-negative.

Stability after a review, blending the long- and short-term updates.

Advance a card's memory state by one review.

Pass Nothing for a card that has never been reviewed; the elapsed time is then ignored and the state is read straight off the initial-stability and initial-difficulty weights.

Fold a whole review history into a memory state.

Each element is (days since the previous review, rating); the elapsed time of the first review is ignored. Nothing for an empty history.

The smallest stability the model will report, 1e-4 days (about nine seconds). This is the s_min the upstream configuration uses for FSRS-7, which is always run with second-precision intervals.

The largest stability the model will report: 36500 days, a century.

1.

10.

One second, expressed in days — the shortest interval nextIntervalDays will return.

A century in days — the longest interval nextIntervalDays will return.