FSRS in practice: what the modern scheduler actually changes

FSRS replaces SM-2's single ease number with a three-part model of memory, and predicts the probability you will recall a card on a given day. In practice it means fewer reviews for the same retention — and it needs data before it can deliver that.

At a glance

The core idea: three numbers, not one

SM-2's whole theory of a card is one ease factor. FSRS uses the DSR model — three quantities per card:

Difficulty. How inherently hard this card is for you. Roughly SM-2's ease, but estimated properly rather than nudged by fixed increments.

Stability. How many days until your recall probability for this card falls to 90%. This is the real innovation. SM-2's "interval" is a schedule; FSRS's stability is a property of your memory that the schedule is derived from.

Retrievability. The probability you would recall the card right now, which decays as time passes since the last review.

The scheduling question then becomes precise. Instead of "multiply by 2.5", it is: given this card's stability, on which day does retrievability fall to the level I asked for? Schedule it then.

Desired retention: the dial that matters

Because FSRS predicts a probability, you can state a target. Set desired retention to 90% and FSRS schedules each card for the day it estimates you have a 10% chance of having forgotten it.

This is genuinely useful, and it is the thing SM-2 cannot offer at all:

• Raise retention (say 95%) and you get shorter gaps, more reviews, and fewer forgotten cards. Sensible before an exam.
• Lower retention (say 85%) and you get longer gaps, far fewer reviews, and more lapses. Sensible for a large collection you are maintaining rather than mastering.

The trade-off is explicit and yours. In SM-2 the same trade-off exists but is buried in constants you would have to hand-tune.

Where the numbers come from

FSRS was trained on a very large body of real review logs — hundreds of millions of reviews from tens of thousands of users. That gives it useful priors before it knows anything about you: a new card starts with parameters that reflect how cards behave in general.

Then it personalises. Anki's optimiser refits the model's parameters against your review history, so the scheduler learns that you find, say, Chinese characters harder than chemistry definitions, and adjusts.

That two-stage design — good priors, then personalisation — is why FSRS beats SM-2 in practice. The commonly cited result is something like a 20–30% reduction in reviews for equivalent retention, which on a large collection is hours a month.

Using it in Anki

Practical notes, since Anki is where most people will meet it:

It is opt-in. Despite plenty of blog posts claiming otherwise, Anki's own manual still describes FSRS as "an alternative to Anki's legacy SuperMemo 2 (SM-2) algorithm". You enable it in deck options.

Enable it, then optimise. Turning FSRS on gives you the general-population parameters. Running the optimiser fits them to your history. Do the second step, or you are getting half the benefit.

Set desired retention deliberately. The default is a reasonable compromise. Do not chase 99% — the review cost rises steeply and the marginal retention is small.

Existing cards are fine. FSRS reads your review history, so a mature collection converts sensibly rather than resetting.

Give it time. Immediately after switching, some intervals will look surprising. The model is usually right and your intuition is calibrated to SM-2.

Why we have not switched yet

We are on SM-2, and the reason is data rather than principle.

FSRS's advantage comes from fitting parameters to review history. A child two months into using flashcards has a few dozen reviews — nowhere near enough to fit anything, and an unfitted FSRS is just a different set of defaults. There is no honest way to claim a 25% efficiency gain for a collection that has no history to learn from.

There is also a proportionality argument. FSRS saves review time, and review time is not what limits a KS3 child: our sessions are capped at twenty cards, roughly six minutes. Cutting six minutes to four solves a problem nobody has. What limits a child is whether the cards exist at all, which is why our effort went into having the tutor write them.

So the plan is deliberate rather than reluctant. The review log records every grade with its timestamp — exactly the shape FSRS needs — so the algorithm can be fitted later without losing history. The trigger for revisiting it is having enough reviews per child for a fit to mean something.

The honest summary

If you are running a large personal collection in Anki, turn FSRS on and run the optimiser. It is a real improvement, it is free, and the people who built it have been unusually rigorous about measuring it.

If you are choosing a system for a twelve-year-old, the scheduler is not the deciding factor. Every algorithm here — SM-2, FSRS, SM-19 — schedules well enough that the difference is measured in minutes per week. The difference that decides whether a child still has a flashcard habit in March is whether the cards ever got written.

FAQ

What is FSRS?

The Free Spaced Repetition Scheduler — an open-source algorithm that models memory as three quantities (difficulty, stability, retrievability) and predicts your probability of recall, rather than multiplying an interval by a fixed ease factor.

Is FSRS better than SM-2?

For a large collection with review history, yes — typically the same retention for meaningfully fewer reviews. For a new collection with little history it has nothing to learn from, so good defaults perform comparably.

Does aitutors.me use FSRS?

Not yet. Our review log was deliberately built so FSRS can be fitted later; the switch becomes worthwhile once there is enough review history to train against.