How review schedules learned to adapt

Imagine sorting paper cards into boxes. After a correct answer, you move a card to a box you check less often. After a missed answer, you put it in a box you check more often.

You've already made the schedule responsive. Today's answer changes what will happen to the card later. You don't need software for that idea.

But the box keeps only a rough summary. Two cards in the same box receive similar treatment even if one answer comes easily and the other takes considerable effort. In this example, the grouping doesn't record that difference.

Keeping a separate record for every card gives a scheduler more to work with. The important change is what information the next decision can use, not merely replacing paper with a screen.

What SM-2 kept

Piotr Wozniak's SM-2 description assigns each item an easiness factor and uses grades for the quality of an answer. The factor helps control how intervals grow and can change with the grade.

The card carries a value that brings earlier responses into later scheduling decisions. It no longer has to follow exactly the same path as every other card.

This is still a compact rule system. An easiness factor isn't itself the probability of remembering the card today. It also isn't the same quantity as difficulty in FSRS. Similar-looking labels can describe different things in different algorithms.

Why the history matters

Imagine two words with the same next review date. One has been recalled successfully many times. The other has repeatedly been missed and was recently learned again. This is an example, not a calculated pair of schedules.

The shared date hides their different histories. Looking only at the date won't tell you which attempts were successful or how much time passed before them.

FSRS, the Free Spaced Repetition Scheduler, stores an estimated state using difficulty and stability. It calculates retrievability, the estimated probability of recall now, from stability, elapsed time, and curve parameters. Retrievability isn't a third independent value stored alongside the other two.

These are quantities in a model, not direct measurements inside the brain. They support two related decisions: estimating access now and choosing a gap for a desired retention target.

An answer must mean something consistent

At the algorithm level, FSRS uses four ratings:

• Again: the answer wasn't recalled correctly.
• Hard: recall succeeded, but with substantial difficulty.
• Good: an ordinary successful answer, with some effort.
• Easy: a successful answer that came without effort.

The Anki manual explains these answer ratings. These are names for outcomes passed to the algorithm, not a promise that every app using FSRS shows four buttons with these labels.

The boundary between Again and Hard matters. If you couldn't recall the answer, choosing Hard reports a success that didn't happen. Even if the revealed answer looks familiar, that recognition isn't the recall the question tested.

The distinction is practical, not moral. The rating describes an encounter with a card, not how disciplined or capable you are. When judging your own answer, ask whether you recalled it before seeing it, then consider how difficult that was.

A forecast can be revised

After a rated answer, FSRS updates the item's estimated state. It can choose a new interval using that state, model parameters, and the chosen recall target. Updating one card's state isn't the same operation as fitting the model's parameters to a person's review history.

A successful answer after a long gap gives evidence of access across that gap. A missed answer gives a different observation. Neither reveals an exact expiration time for a memory, and one failure doesn't prove that the whole model is wrong.

This history is a comparison of scheduling ideas, not a claim that every newer method directly descends from the one before it. The shared practical question is how much the next date should learn from the last answer.

With FSRS, understanding the answer starts with three terms: difficulty, stability, and retrievability.