The scheduler

The decks use Anki’s own scheduler, SM-2 as Anki’s v2 and v3 schedulers run it, so a habit formed in Anki carries over, and a number you know from Anki means the same here. This page gives the rules with the deck’s default options; every number in italics below is one of those options.

1. Four states#

A deck’s counts join learning and relearning: both come back within minutes, and both are red on the study page.

2. Learning: the first minutes#

A new exercise goes through the learning steps — by default 1 minute, then 10 minutes:

Again

sends it back to the first step: it returns in a minute.

Hard

repeats the step it is on. On the first step it waits halfway between the first two steps — 5½ minutes, shown as 6m. (With a single step, Hard waits half as long again as that step.)

Good

moves it to the next step; Good on the last step graduates it to review, due in 1 day (the graduating interval).

Easy

graduates it at once, due in 4 days (the easy interval) — and always at least a day later than Good would have.

With no learning steps at all, an exercise graduates the first time you press Good.

3. Review: the waits grow#

A review waits whole days, and each exercise carries its own ease, a multiplier that starts at 2.5 when it graduates:

• Good multiplies the interval by the ease: 4 days become 10.
• Hard multiplies it by 1.2 only, and lowers the ease by 0.15.
• Easy multiplies it by the ease and by 1.3 (the easy bonus), and raises the ease by 0.15.
• The three are kept apart: Hard always gives at least one day more than the interval had, Good at least one day more than Hard, Easy at least one day more than Good.
• Answered late, an exercise gets credit for the days it waited beyond its due date: half of them on Good, all of them on Easy. Knowing a thing after a longer wait than planned says it is better known than planned.
• No interval is longer than 36,500 days (the maximum interval) or shorter than 1 day.

An exercise you keep knowing thus comes back after 1 day, 3 or 4, about 10, about 25, two months, five… — and one you find hard sees its ease fall, and its waits grow more slowly.

4. Lapses: when you forget#

Again on a review is a lapse. The ease drops by 0.20 — never below 1.3, the minimum ease — and the exercise relearns: it comes back after the 10-minute relearning step. Good on that step puts it back into review, due in a day; Easy, a day later than that; Again, Hard and Good on the step work as in learning.

The deck counts the lapses: the list on the deck’s page shows them beside the reviews, 4 reviews · 1 lapse.

5. The spread#

Intervals of three days and more are moved by a few percent — up to 15% for a week or less, 10% up to twenty days, 5% beyond, and always at least a day — so that forty exercises added on the same evening do not all fall due on the same morning.

Unlike Anki, the spread is not drawn by lot. It is worked out from the exercise itself, the button and how many times it has been answered, so the label on each rating button is exactly the interval the answer will get.

6. The day starts at four#

A deck’s day begins at 04:00, local time, as Anki’s does. An answer given at one in the morning counts for the evening before, and an exercise due “tomorrow” is due from four o’clock tomorrow morning — for the whole of that day. The change to and from summer time moves nothing: a review due on a day stays due on that day.

7. Daily limits#

At most 20 new exercises and 200 reviews a day, counted from the answers already given that day. What a limit holds back waits for the next day — the decks page then says next exercise tomorrow — and learning steps are never limited, since holding back an exercise you are in the middle of learning would undo the learning. A limit set to 0 holds its exercises back every day: a deck with 0 new exercises a day studies only what it already knows.

8. The two sides of a pair#

The two cards a both (repeat) flashcard makes (in a deck) are linked, and the scheduler will not ask you one right after the other: once you have answered one, the other is held back until the next day starts — Anki’s bury siblings. The hold is read from the answers already given, so nothing is stored for it; it ends by itself at four in the morning, and at once if the other card is deleted, unlinked or set to new. A limit above still counts only what was shown, and cram ignores the hold, as it ignores the schedule.

9. What comes next#

Of everything due, the study page shows first a learning step that is due, then a review that is due — the one that has waited longest first — then a new exercise, in the order they were added. When none of those is left today, a learning step due within the next 20 minutes is shown at once rather than making you wait: Anki calls that learning ahead.

10. A worked example#

One exercise, answered each time on the day it fell due, with the deck’s default options. The last four columns are what the four buttons offered for the next answer — what you would have seen on the study page.

The spread makes another exercise’s numbers differ by a day here and there; the shape is the same. Note the lapse: after 24 days the exercise starts again from a day — the new interval option decides how much of the old interval a lapse keeps, and by default it keeps none — but with an ease a little lower, so it climbs more carefully the second time.

11. For those who know Anki#

The deck keeps Anki’s defaults and its rules, with four small differences:

• the spread is computed, not random, so the labels are exact;
• halves round up (Anki rounds its own way);
• the minimum interval holds for every interval, not only after a lapse;
• a button’s label for a day interval reads the interval (1d), not the hours left until the next four o’clock.

The scheduler is markdown/app/srs.py, and it keeps every answer: each exercise’s schedule/<id>.json holds its state and the whole list of your answers, with the rating, whether the check found you right, and the interval and ease that followed.