Ask most spaced-repetition tools when to show you a card again, and the honest answer is: a table. A fixed sequence of intervals - a day, three days, a week, two weeks - applied the same way to every card, regardless of what that specific card is or how you've actually done on it before.
FSRS takes a different approach. Instead of a table, it's a model of your memory for that specific card, updated every time you review it.
In short
- FSRS stands for Free Spaced Repetition Scheduler: an open-source algorithm developed by Jarrett Ye with a community of contributors.
- It describes each card with three numbers - difficulty, stability and retrievability - and updates them after every review.
- Because it predicts the chance you will remember a card right now, it can schedule to a target you choose instead of following a fixed interval ladder.
- The algorithm has parameters. Defaults are an average over many learners; fitting them to your own history makes the schedule fit you.
- Cirello uses FSRS-6 and retrains its parameters nightly.
Where FSRS comes from
FSRS is an independent open-source project driven by its community. Its algorithm is credited mainly to Jarrett Ye, and the project documents it in public (the ABC of FSRS). The name is literal: a scheduler, free to use and inspect, whose job is to answer one question for every card on every day - should this be shown now?
It has gone through several versions over time. FSRS-6, the one Cirello uses, is the most recent, with 21 parameters in the formulas that describe how a card is remembered. The project notes that, unlike earlier versions, FSRS-6 uses all of the reviews, not only the first one in each day.
What FSRS actually models
Where a fixed-interval system asks one question - "how long since last time?" - FSRS tracks three separate properties per card. Together they're called the DSR model:
- Retrievability is the probability that you can recall the card at this moment. It falls as time passes since your last review.
- Stability is how long the memory lasts. Precisely, it's the number of days it takes retrievability to fall from 100% to 90%. A card with a stability of 30 days is one you'd have a 90% chance of recalling a month after reviewing it.
- Difficulty is how inherently hard the card is for you. It influences how quickly stability grows each time you review successfully.
Grade a review "easy" and the model shifts its estimate of that card's difficulty down; grade it "again" and it shifts up, then schedules a tighter interval to compensate. After a successful review, stability goes up - and it goes up by more when you recalled the card at a moment of low retrievability, because a hard-won recall strengthens memory more than an easy one.
The practical effect: two people studying the same deck end up with two different review schedules, because the model is fit to what's actually happening in each of their sessions, not to a table both of them share.
A worked example, without the maths
Take two cards, both new this morning. One is a common word you already half-knew; the other is a tangled rule you always get wrong.
The word gets graded "easy" on the first review. The model estimates low difficulty and a relatively high stability, so the next review lands several days out. When it comes up again and you recall it without effort, stability grows again and the next gap stretches further. Within a few reviews the card is on a schedule of weeks.
The rule gets graded "again". Difficulty is high, stability is low, and the card returns soon - and keeps returning, at first in short gaps. Each success still pushes stability up, but more slowly than for the easy card. In a few weeks it may be scheduled at a similar gap to the first card, because the model has watched you actually learn it.
Nothing about this required you to sort the cards into "easy" and "hard". You just answered honestly, and the numbers followed.
Why a model beats a table
A table has to guess in advance. It can't know which cards are hard for you, and it can't adapt when your memory turns out to be better or worse than average. Some tables adjust with a per-card ease value, and that is a real improvement - FSRS vs SM-2 compares the two ideas directly. But an ease number is a single dial. FSRS separates how hard a card is from how stable your memory of it has become, and it models forgetting as a curve over time rather than a fixed step.
The FSRS project reports, from simulations, that it needs noticeably fewer reviews to reach the same retention as the older method, even with default parameters. That is the project's claim about its own tool, not a guarantee for any individual, but the direction is consistent with the design: if the model predicts forgetting better, fewer reviews are wasted on cards you would have remembered anyway.
Why "nightly-retrained" matters
Most systems that use this style of model still ship it with parameters fixed at release - tuned once, on someone else's review data, then frozen. Cirello retrains FSRS-6's parameters every night, on the review history that's actually accumulated.
That means the model isn't just personalized per card - it's continuously re-fit as more of your review history exists to learn from. Early on, with little data, it leans on reasonable defaults. As you review more, the schedule increasingly reflects your own memory, not an average of everyone else's.
The schedule gets more accurate the more you use it - not because you configured anything, but because there's more of your own data to fit it to.
The mechanics - what runs, when, and why nightly - are in inside Cirello's FSRS-6 pipeline.
The target FSRS aims at
Because FSRS estimates the probability you'll recall a card, a scheduler can ask: when will that probability drop to a chosen level? That level is called desired retention. Aim higher and cards come back sooner and more often; aim lower and you review less but forget more. It's the main trade-off in any FSRS-based scheduler, and it has its own explainer: desired retention in FSRS.
What this changes day to day
You won't see FSRS-6 directly - there's no dial to turn, no setting to tune. What you'll notice instead is what it produces: cards you find easy drift toward longer gaps between reviews, cards you consistently struggle with come back sooner and more often, and the review queue on any given day stays close to what's actually due - not padded with cards a fixed table thinks you should see regardless of how you've been doing.
Backlogs also behave better. If a rough week leaves you with overdue cards, the model knows how long each has actually been since you saw it and adjusts, rather than treating every overdue card the same. That's part of why a calmer way to study dense material leans on this kind of scheduling.
Questions people ask
Does it work from the first review? Yes, in the sense that it needs no history to start: the defaults give a reasonable first schedule. What improves with time is the fit. After a few weeks of reviews the schedule has real evidence about how you forget; after months it has a lot.
Do I need to understand the maths? No. The three ideas - difficulty, stability, retrievability - are enough to read most discussions, and using the scheduler requires only that you answer and grade honestly.
What if I grade inconsistently? The model treats your grades as the truth. If you call things "easy" that you barely recalled, it will space them too far apart and you will miss them more often. A simple habit helps: grade what happened, not what you wish had happened, and when in doubt grade the lower option.
Is a more complex model always better? Not automatically. A model earns its place only if it predicts your forgetting better than a simpler one would, and that is why fitting it on real review history matters more than the formulas themselves.
What FSRS doesn't do
A scheduler only decides when. It doesn't make a badly written card good, and it can't tell whether you looked at the answer before honestly trying to recall it. The model is only as good as the grades you give it. Writing small, clear cards matters (how to write good flashcards), and so does answering honestly. If you're new to the underlying idea, begin with what spaced repetition is and why it works.
That's the whole point of using a model instead of a table: the schedule adapts to you, instead of asking you to adapt to it. To see how it fits into a complete app, visit the FSRS app page.