← Blog

How often should you review a word you already know?

There is a real answer to how long you should wait before reviewing a word. Then you do the division.

Every few weeks I go back through the words Kezi says I have mastered and read down the list. Most of them are fine. Then there is one I do not recognise, and I take it out of the mastered pile and put it back into practice.

翻译 has been through that cycle more than once. I have learned the word for "to translate", lost it, learned it again, and lost it again.

What I could never work out is how often I should be doing this. Every week felt like too much. Every couple of months felt like I was letting things quietly rot. I had no way to decide, so I went looking for whether anyone actually knew.

A while ago I wrote about why the words you got right on Monday are gone by the following week. The short version is that getting a word right once and never being asked again is the condition that collapses. This post starts where that one ends, with the obvious next question.

So: how often?

The research has an answer, and it is a good one

There is a whole culture of argument about review schedules. Most flashcard apps are built on expanding intervals, where each successful review pushes the next one further out: three days, then ten, then a month, then three months. People tune these ladders the way other people tune carburettors.

The ladder is close to irrelevant.

Karpicke and Roediger tested expanding intervals against evenly spaced ones three separate times. Their conclusion, in their own words: "Expanding the interval between repeated tests had little effect on long-term retention in 3 experiments. The important factor for promoting long-term retention is delaying initial retrieval to make it more difficult."

Not the shape of the ladder. The waiting.

Which leaves the real question: how long is long enough? Cepeda and colleagues ran that one properly, with 1,354 people learning obscure facts and being tested after 7, 35, 70 or 350 days. For each horizon they looked for the gap that produced the best recall.

The answers were 1 day, 11 days, 21 days… and then 21 days again.

How long to wait before asking again

The gap that produced the best recall, for four different horizons. Cepeda et al. (2008), 1,354 participants.

The gap grows with how long you want to remember something, and then it stops. Somewhere around three weeks the curve flattens and stays flat, whether you are trying to remember a word for ten weeks or for a year.

That is a genuinely useful finding. Three weeks, roughly, and stop fiddling with the ladder.

That lasted until I counted.

Then you do the division

Here is the part nobody selling you a review algorithm will mention.

A review session that stays a two-minute thing holds about a dozen words. That is not a Kezi limitation, it is what two minutes is. Anything longer stops being something you do while the kettle boils, and the thing you do while the kettle boils is the only kind of practice that survives a busy week.

So the number of words you can be asked about each week is roughly fixed. The number of words you have mastered only grows. Divide the second by the first and you get your real review interval, whatever any app promises you.

What you actually get

A two-minute session holds about a dozen words. Divide your mastered words by that, and this is the gap you really have. The horizontal line is the three weeks the research recommends.

Past roughly three dozen mastered words, the schedule stops being a scheduling decision. Your interval is set by division. And every learner who sticks with this for a few months is well past three dozen words, which means the interval argument is settled by arithmetic long before it is settled by science.

No app can hand you the optimal gap. Not Kezi, not the one with the clever algorithm, not the one you have been feeling guilty about abandoning. The only thing any of them can actually decide is which words you see first.

So why bother at all

Because the gap between a good interval and a mediocre one is small, and the gap between being asked and not being asked is enormous.

That is the finding from the first post, and it is worth repeating exactly once. In the 2008 study, words that kept being tested after the first correct recall sat at around 80% a week later. Words that were dropped from testing sat at 36% and 33%, including the ones people kept dutifully rereading.

A word reviewed at twenty weeks is enormously better off than a word reviewed never. Late is a rounding error. Never is the whole problem.

Which turns the question from scheduling into priority. If you cannot get to everything, and you cannot, then what should come back first?

The oldest thing you still think you know.

Not the words you feel shaky about, because that feeling carries no information, everyone in that study felt fine about all of it. The ones you have not been asked about in the longest time. That is the only signal that means anything, and it is the one your notebook cannot give you and your memory will not.

I stopped trying to answer "how often should I review?" It has no answer. The question I can answer is "what have I not been asked about in ages?", and that turns out to be the same question wearing more honest clothes.

Kezi's job, then, is not to promise you a perfect interval. It is to hand back your oldest words, a few at a time, in something small enough that you finish it. Not a pile of due cards that grows while you sleep, because a pile you cannot clear is how people quit.

翻译 is currently in my practice pile. It will probably come back. I have stopped treating that as a failure and started treating it as the system doing what it is for.