“Spaced repetition” has become a slogan. It is also a real finding, and the real finding is less tidy than the apps make it sound.
If you meet something twice, the time between those two meetings changes what you still have later. A gap of zero (study, then study again at once) is a poor way to keep a fact for a month or a year. A gap that is too long is also worse than a middle one. The middle that works is not a constant. It moves.
That is the whole useful sentence. Everything below is so we do not turn it into a clock we never measured.
The big review (2006)
Cepeda, N. J., Pashler, H., Vul, E., Wixted, J. T., and Rohrer, D. (2006). Distributed practice in verbal recall tasks: A review and quantitative synthesis. Psychological Bulletin, 132(3), 354–380.
They pulled the verbal-recall spacing literature together: 317 experiments, 839 assessments, from more than 400 reports (only a thin slice of those reports had looked at delays as long as a day). The regular result: space beats mass. The gap that does the most good gets longer as the time you want to remember gets longer.
They could not, from that literature, tell a teacher “use 3 days” and be done. Most of the old experiments were minutes to hours. That is why the same lab then ran a study that lasted up to a year.
The year-long map (2008)
Cepeda, N. J., Vul, E., Rohrer, D., Wixted, J. T., and Pashler, H. (2008). Spacing effects in learning: A temporal ridgeline of optimal retention. Psychological Science, 19(11), 1095–1102.
1,354 people, mean age 34 (range 18–72), learned 32 obscure true facts. Session 1: train to one perfect recall per fact, with feedback. Then a gap of 0 to 105 days. Session 2: the same list twice, answer shown after each item, study time held even across groups. Then a retention interval of 7, 35, 70, or 350 days. Final tests: recall, then five-choice recognition, no feedback.
Twenty-six combinations of gap and delay.
For every delay they tested, recall first rose as the gap grew, then fell. An inverted U, not a straight “longer is better.”
Optimal gaps among the ones they ran (recall): 1 day for a 7-day test, 11 days for a 35-day test, 21 days for a 70-day test, 21 days for a 350-day test. After they interpolated the curves: about 3, 8, 12, and 27 days. As a share of the delay, that is roughly 43%, 23%, 17%, and 8%. The abstract’s plain-language version: about 20–40% of a one-week delay, down to about 5–10% of a one-year delay. In their fitted surface, the best gap for a 350-day test sat at 23 days, about 7%.
Compared with a zero-day gap (study the facts, then immediately study them again), the best gap they found improved recall by 10%, 59%, 111%, and 77% at the 7-, 35-, 70-, and 350-day tests. Averaged across those delays, the best gap versus zero was a 64% lift in recall (d = 1.1).
Read that again if you cram. The zero gap is the night-before twice. It was the worst of the gaps they plotted, once you care about next month.
What this paper does not prove
- It does not give MnemoBooks, or Anki, a blessed interval. These were trivia facts trained to criterion, then restudied on a fixed schedule, then tested once. A bilingual story book is a different task.
- It does not say “the algorithm found your exact second.” The useful gap is a range that moves with the deadline.
- It does not say a longer gap is always better. Past the peak, performance dropped.
- It does not settle expanding vs equal spacing. That is a different argument. This paper maps two meetings and one later test.
What to do with it
Decide when you want the word to still be there. If the answer is “Sunday,” a restudy tomorrow beats a restudy tonight. If the answer is “next year,” a restudy in a few weeks beats a restudy tonight, and also beats waiting until the week before.
You will feel worse at the second meeting if you waited. That worse feeling is often the point. The 2008 students who restudied after a real gap had forgotten more at that restudy. They still won the test that sat months later.
Where a book fits
A MnemoBooks title returns the same 1,111 words across 85 short stories, then asks for them again in the exercises. That is spacing plus a later retrieval, not a claim that we timed each word to Cepeda’s ridgeline. We did not. Anyone who tells you a printed book hit “the scientifically optimal interval” is selling a number the 2008 paper does not contain.
What we can say, without reaching: you will meet the core again after you have had a chance to start losing it, and you will have to produce it, not only see it. That is the shape the literature likes. The catalog is 22 languages. More on the method.
Sources: Cepeda et al. (2006), Psychological Bulletin, 132, 354–380; Cepeda et al. (2008), Psychological Science, 19, 1095–1102.
For the wider toolbox, and what actually exists per language, see how people actually learn languages.