The number to think about is not the interval, it is how many new items you add each day. An interval is a consequence of your own answers and the program's schedule; a new card is a decision, and it commits you to a review tomorrow, one in three days, one next week, one next month and one after that.
Add fifty a day for a fortnight and the reviews arriving in six weeks are not fifty a day. They are the accumulated tail of everything added since, and the number is large enough that most people stop, which is why abandoned decks are almost always decks that were filled quickly.
Eight rules for making a system work
- review just before you forget, which is the entire principle. Everything below is an implementation of it.
- the first review is the same day the card is made, and skipping it costs more than any later interval.
- one day, three days, a week, a month is a workable manual schedule for anybody not using software.
- twenty new cards a day is a ceiling for most people. Each new card generates around ten future reviews, so the daily load compounds quickly.
- reviews compound and new cards do not. Adding fifty cards on a keen Sunday produces an unmanageable Tuesday three weeks later.
- never skip a day of reviews. Two skipped days produce a backlog that most people never clear, and the backlog is the commonest reason people abandon the method.
- delete cards you keep failing. A card failed six times is badly made, not badly learned, and it should be rewritten or removed.
- the algorithm schedules, you still study. The software decides when; it does nothing about whether the card was worth making.
The intervals grow, and the growth is the whole design
The gaps are not evenly spaced and they are not meant to be. Roughly: the same day, then a few days, then a couple of weeks, then a month or two, then longer. Each successful review pushes the next one further out, so a word that is holding costs you almost nothing after the first month.
That is the return on the whole system. A hundred items reviewed daily is unmanageable and a hundred items on expanding intervals is a few minutes, and the difference is entirely in the shape of the gaps rather than in any individual review.
It also means the early days are the expensive ones. A new card is at its most demanding in its first week and nearly free by its third month, so the cost of an addition is front-loaded and the benefit is not, which is exactly the shape that makes people over-add.
Rate the retrieval, not the feeling
The only real input you give is the button you press after each card, and most people press it on how the card felt rather than on what happened. A card that came back correctly after a moment's effort is not a hard card, it is the system working, and marking it hard drags the interval back and produces more work for a card that did not need it.
The rule that keeps a schedule honest is narrow: did the answer arrive, or did you look. If it arrived, it was a success, however uncomfortable it felt. Effort is not a symptom of failure here; it is the thing that made the review worth doing.
Where it stops: adding faster than the future can absorb
The stall is a backlog, and it is built in the first three weeks by somebody enthusiastic. New cards are cheap to add and the reviews they generate arrive later, so the feedback that would have slowed you down is delayed by exactly the period in which the damage is done.
What it looks like a month later is a daily count that will not come down whatever you do, a session that takes an hour, and a growing sense that the method is punishing. None of that is the method. It is an intake decision made a month earlier by somebody who could not see the bill.
The way out of an existing backlog is unpleasant and short: set new cards to zero and keep reviewing until the count falls, which takes a few weeks and works reliably. Deleting the backlog also works and costs you the cards.
Looking at next month before you add today
Anki shows a forecast of upcoming reviews, free with the program, and it is the screen that should govern how many new cards you add. It plots the reviews already committed for the coming weeks, so before adding thirty items you can see what the last thirty are going to cost you in a fortnight.
It replaced discovering the workload by living through it, which is how everybody learns this and which costs a month.
Its limitation is that it is a projection built on cards behaving as expected, so it under-counts: every card you fail comes back sooner than the forecast assumed. Treat the line as a floor rather than an estimate.
What it looks like when it goes wrong
The recognisable failure is a review count that grows on days you did a full session. That means failures are being generated faster than successes, which is almost always intake rather than difficulty.
The second is a session spent mostly on a handful of items that keep failing. Those are not hard words; they are usually badly built cards, and grinding them is the most expensive thing in the deck.
What the schedule is exploiting, and why a review pays most when it was nearly forgotten, is why you forget, and the schedule that beats forgetting, and what a card has to look like for any of this to work is making cards that teach rather than cards that pile up.
The question underneath the intake decision is where the items are coming from at all, because cards taken from a list generate the same workload as cards taken from your own reading and are much easier to abandon: filling your day with the language without moving country.