What changed in FSRS-6

You answer a word correctly, then see a new review date. Upgrading the scheduler may change that date even though your answer hasn't changed. What did the new version learn to describe differently?

FSRS-6 changes two parts of the model: how predicted recall declines with time, and how closely spaced reviews update stability. The goal stays familiar: estimate the state of an item, then calculate a useful next review time.

A curve whose decay can be fitted

FSRS-6 has 21 model parameters, compared with 19 in FSRS-5. One of the new parameters controls the decay of the forgetting curve. These are coefficients used by the model, not 21 separate scores stored for each word.

When parameters are optimized against review history, that decay can be fitted rather than held to the previous fixed value. A scheduler using default parameters still uses a specified decay value; it doesn't independently retrain the curve after every answer.

Stability still has the same reference point: the interval at which predicted recall reaches 90%. Making decay trainable doesn't turn stability into a different kind of measurement.

The official algorithm specification describes these changes. A trainable parameter is an opportunity to fit a model, not proof that a particular app already fits it separately for every user.

Closely spaced reviews update stability differently

FSRS-5 already had a rule for short, same-day reviews. FSRS-6 adds a dependence on the item's existing stability. With the other inputs fixed, the relative growth factor is larger when stability is small and smaller when it is large.

Think of a new word encountered again soon after learning it, compared with a familiar word reviewed again just as quickly. The revised rule gives the model a way to treat those histories differently. This is an explanation of the formula, not a claim that repeating either word immediately is the best study choice.

The app still decides which practice events count as scheduling evidence. Modeling a short interval and filtering an accidental duplicate are separate responsibilities.

What did not start with version 6

Difficulty updates with linear damping already appear in the FSRS-5 specification. That rule reduces the size of the difficulty adjustment near its upper limit. It belongs in a description of the model, but isn't a new FSRS-6 invention.

That distinction may seem small when you're only looking for tomorrow's review date. It matters when explaining why a version changed. More parameters alone don't establish better learning outcomes, and a feature should be credited to the version in which it appears.

Updating an item and fitting a model

An ordinary review updates the difficulty and stability associated with an item. Parameter optimization is a separate process that looks across review history to fit the rules used for those updates.

An app can therefore adapt each item's schedule while using the same default parameter set for many learners. Later, it may fit a user's parameters and keep updating individual items with that fitted model.

For you, the distinction is practical. A changing review date can reflect an updated estimate for one word. It doesn't, by itself, mean the app has retrained its entire model from your history.

OOWN uses FSRS-6 in its scheduling implementation. Its exercises, however, don't all look like four rating buttons. The final step in this series is to follow how an exercise result becomes a signal the scheduler can use.