Every scheduling model has to answer the same question eventually: whose data trained it? Most ship an answer once - a set of parameters tuned before release, frozen in place, applied to everyone the same way from day one. Cirello's answer is different: FSRS-6's parameters are retrained every night, against the review history that's actually accumulated.
What "nightly retrained" means mechanically
FSRS-6 estimates, per card, how difficult it is and how fast your recall of it decays. Those estimates start from reasonable defaults, then get refit as review data comes in. The retraining job runs nightly rather than continuously for a simple reason: a single review or two shouldn't be enough to swing the model - it needs a meaningful batch of fresh outcomes to refit against, and a nightly cadence is that batch, applied consistently instead of firing off unpredictably.
Practically, that means the model you're scheduled against tonight is a slightly better fit to your actual memory than the one you were scheduled against last night - not because anything you configured changed, but because there's more of your own review history behind it.
Why this matters more the longer you use it
A model trained once and shipped frozen degrades quietly. It doesn't get worse at scheduling in any dramatic way - it just stays exactly as accurate as it was on day one, forever, no matter how much you personally review. A model retrained nightly does the opposite: accuracy compounds. Ten reviews in, it's working off thin data and leaning on defaults. A few hundred reviews in, it has real signal on how you specifically perform - which cards you find easy that most people don't, which ones take you longer than average, how fast your recall actually decays for content in general versus for one stubborn topic in particular.
The schedule isn't tuned once for everyone. It's re-fit every night, against your own review history specifically.
What you actually experience
You won't see a dashboard for this - there's no setting to adjust, no number that goes up. What you'll notice instead is indirect: review sessions that feel proportionate to how well you actually know the material, cards you've nailed repeatedly drifting toward longer and longer gaps, and cards that keep tripping you up showing up again sooner, without you ever telling the app which category a card belongs in.
That's the entire design goal of retraining nightly instead of once: the scheduling gets quietly better at predicting you, specifically, the more you use it - not because the algorithm changed, but because it never stopped learning from what you actually did.
If you want the foundations, what FSRS is explains the memory model being fit, FSRS vs SM-2 shows how it differs from older schedulers, and desired retention in FSRS covers the target the schedule aims at. The app-level summary is on the FSRS app page.