How the schedule decides when to ask

Most flashcard software asks you about a word on a ladder. Right answer, next rung. Wrong answer, back to the bottom. The rungs are fixed in advance and they are the same for everybody, which means they are wrong for everybody, just by different amounts.

The alternative is to model the memory itself, and then place the review where the model says it is needed. That is what a modern scheduler does, and the one in here is FSRS.

Two numbers per card

• Stability is how many days your memory of this card lasts. It is not how long ago you saw it, it is how durable the memory currently is.
• Difficulty is how hard this particular card is for you. Not in general, for you. The same word can be trivial for one person and a nightmare for the next, and a shared difficulty rating throws that away.

From those two, plus the days since you last saw it, the scheduler works out a third number: retrievability, the probability you could pull the word up right now. It falls off in a curve, quickly at first and then more slowly, which is the shape Ebbinghaus described in 1885 and which has survived a century of people trying to find a better one.

Where the review goes

The review is placed at the moment retrievability crosses your target. The default target is ninety percent, which means: ask me while I can still just about get it.

That number is the whole tuning dial. Set it higher and you see everything more often and forget less, at the cost of a lot more repetitions. Set it lower and you do less work and lose more. Ninety is a good default because the curve is steep below it, so the extra work of going to ninety five buys much less than it costs.

The two failure modes either side are worth naming, because they are what a fixed ladder does to you constantly. Ask too early and the repetition is nearly free, which sounds good and is actually waste, because a memory you were in no danger of losing gets no stronger for being tested. Ask too late and the word is gone, and you are not reviewing, you are learning it again from the beginning.

What a good answer buys you

Answer a card well and its stability grows, so the next gap is longer. Keep answering it well and the gaps stretch out fast. Seven consecutive good answers on one card, with the weights that ship here, takes you from ten minutes to a little over three years.

Miss it and stability collapses, but only for that card. That is the part a class-wide review schedule cannot do. When thirty students share one review queue, one student forgetting one word either drags everybody back or is ignored, and both are wrong.

Learning steps, before any of this

A brand new card is not ready for a curve, because there is no memory yet to model. So the first two exposures are on fixed short steps, ten minutes and a day, exactly as Anki does it. Only once the card has been answered well from that does it graduate into the scheduler proper.

The honest limitation

The weights that ship are the published ones, fitted across a very large population. They are not fitted to you. Personal fitting needs a few thousand of your own reviews before it beats the general weights, so it is a later feature, and the columns to hold your own weights are already in the table waiting for it.