Personalized spaced repetition

Words fade from memory in predictable ways, and that can be put to work. BrioVocab schedules reviews with FSRS, one of the most modern spaced repetition algorithms, and tunes it to how you personally memorize. Below is how it works, without formulas. Click a reference to see the publication right on the page.

1. Forgetting can be predicted

The curve drops fast, but every review makes it flatter

In 1885 the psychologist Hermann Ebbinghaus memorized nonsense syllables and recorded how quickly they slipped away (). The result is the forgetting curve: steep in the first hours, then slower and slower. Modern replications of the experiment show the same picture (). The useful part: a successful review slows forgetting down, so the next encounter with a word can be pushed further out. That is the whole method: review at growing intervals, shortly before the word slips away (). Review too often and you waste time; too rarely and you relearn from scratch.

forgetting curve · try it

no reviewswith reviews

without reviews, memory of a word melts in a couple of weeks

2. Several generations of algorithms

From a cardboard card box to a model trained on millions of reviews

The method is older than computers. In the 1970s Sebastian Leitner sorted paper flashcards into boxes: the better you know a card, the less often you open its box (). Computer schedulers of the SuperMemo family arrived in the late 1980s (); their early SM-2 scheme still runs under the hood of many apps. The next step came from models trained on real data: that is how Duolingo predicted its users' forgetting (). FSRS appeared in 2022, an open algorithm trained on hundreds of millions of real answers (). In open scheduler comparisons it consistently ranks among the most accurate. BrioVocab runs its sixth generation, FSRS-6, and our implementation is checked against the reference one.

how the algorithms evolved

1885 the forgetting curve
1972 Leitner card boxes
1987 SuperMemo, SM-2
2016 models trained on big data
2022 FSRS
now FSRS-6 in BrioVocab

3. What the algorithm knows about each word

Not one queue for everyone, but three measures per word

FSRS keeps three measures for every word you study. Stability tells how slowly the word fades: the higher it is, the longer the pause before the next review. Difficulty reflects how hard the word is for you specifically: intervals for difficult words grow more cautiously. The third measure, the probability of recall, melts away with each day since the last encounter. Researchers proposed splitting memory into stability and momentary availability long before FSRS (). As soon as the probability drops toward the threshold, the word joins the review queue. After your answer, stability and difficulty are recalculated, and the word gets a new date.

probability of recall · drag it

la plagebeach

95%review scheduled for day 6

days since the last review: 3

stability of this word: 6 days · difficulty: medium

4. A personal memory model

The algorithm retrains on your answers every day

FSRS-6 behaviour is set by 21 parameters. A new user starts with the standard values: they come from large datasets and work well on average. Memory differs from person to person though, so for subscribers BrioVocab recalculates all the parameters daily from their personal answer history. It is done carefully: random slips are filtered out, recent answers weigh more than old ones, and while the history is small the model changes cautiously instead of drawing big conclusions from a couple of mistakes. As a result, your intervals are planned by your own model, not a shared one. Few spaced repetition apps go that far.

intervals of one word · flip it

starting model

J+1J+3J+8J+18

personal weights

the model is recalculated every day from your answers

5. You stay in control of reviews

The algorithm listens to every answer, but the final say is yours

Grades reach the algorithm from every kind of training: Super Coach as well as self-graded flashcards. On the cards the forecast is shown up front, right on the buttons: how many days the word will stay away for each possible answer. There are manual controls too. Known this word since school? Mark it as known and the trainings stop showing it. Changed your mind, or noticed it faded after all? One button puts it back into the active queue.

flashcards · self-grading · click one

l'oragethunderstorm

the forecast is computed by the model for this very word

le pontin training

6. Progress you can see

Nine levels and protection from random dips

Inside the algorithm a word's stability is measured in days and changes smoothly. To make it easy to follow, BrioVocab turns it into a scale of nine levels, from a new word to a reliably learned one. The scale is protected from random jumps: a word only drops a level when its stability has sagged noticeably. One bad answer does not wipe out weeks of work. Next to the scale sits the day's summary: how many words are due, how many new ones are coming, and how many are already learned.

word levels · L0–L8

L0L1L2L3L4L5voilierL6L7L8

due today · 12new · 5learned · 348

Try it on your own memory

Words at the right moment, not on a shared schedule

Start for free: training in BrioVocab is already scheduled by FSRS. A subscription adds a personal memory model, recalculated every day.

Start for free