Spaced repetition: how it works, how to do it by hand, and where it breaks
You already know the opposite of spaced repetition. It is cramming: reading the same material six times the night before, walking into the exam with it fresh, and finding a month later that almost none of it is there. Cramming works for the morning after. It does nothing for the year after, and the reason it does nothing is the same reason spacing works.
Spaced repetition is the practice of reviewing something at increasing intervals: a day after you learn it, then a few days, then a couple of weeks, then months. Each review comes just before you would have forgotten, and each successful recall pushes the next one further out. The idea is a century old, the evidence for it is unusually solid by the standards of education research, and the software that runs it has been around since the late 1980s.
This is what the effect is, why the gap is the point, how a schedule decides when something comes back, how to run one with a shoebox and index cards, and the two ways it fails for nearly everyone who tries it.
What the spacing effect is
Learn something once, then review it. If the review happens immediately, it adds little; you still have the thing and the review is a repetition of a memory that has not yet needed strengthening. If the review happens after a gap, when the memory has started to fade and recalling it takes effort, the same review adds a great deal. Two reviews separated by time beat two reviews back to back, and the advantage grows with how long you want to remember.
That is the spacing effect. Ebbinghaus noticed it in his own syllable experiments in the 1880s; a meta-analysis by Cepeda and colleagues in 2006, covering more than two hundred studies, found the advantage of spaced over massed review to be one of the most reliable results in the field. Dunlosky's 2013 review of study techniques rated distributed practice as high utility, alongside practice testing and above everything else.
The mechanism, as far as it is understood, is the one described in the active recall article: a retrieval that takes effort strengthens the memory more than one that comes easily, and the effort is greatest just before the memory would have gone. Spacing is a way of arranging your reviews so that each one lands in that window.
Why the gap matters more than the count
The intuition most people bring is that more reviews are better. They are, but the size of the gap matters more than the number of reviews, and the right gap is not a constant.
Cepeda and colleagues, in a 2008 study, taught people facts and then varied both the gap between the first and second study session and the delay before the final test. The best gap depended on the delay: to remember something for a week, a gap of a day or so was best; to remember it for a year, a gap of several weeks was best. Too short a gap and the second review was wasted on a memory that did not need it. Too long and the memory was gone before the review could rescue it.
For a reader rather than an experimental subject, the practical consequence is that the schedule has to grow. The first review comes soon, because the first drop is steep. Each successful review then earns a longer gap, because the memory is now more durable and the next window of useful effort is further away. A card you have recalled correctly five times over four months might not need to come back for a year. A card you have missed twice this week needs to come back tomorrow.
Successful recall lets the gap before the next review grow.
How a schedule decides
A spaced repetition system is a scheduler. You tell it how a review went, and it decides when the card comes back.
The oldest working design is the Leitner box, from the 1970s: a set of compartments, cards move up one when recalled and back to the first when missed, and each compartment is reviewed less often than the one before. It is a schedule you can hold in your hands and it is still worth knowing, because every later system is a refinement of the same idea.
The first software refinement was SM-2, published by Piotr Wozniak in the late 1980s for SuperMemo and later adopted by Anki. Instead of fixed compartments, each card carries a number that grows with each successful review and shrinks with each miss, and the next interval is computed from it. It ran unchanged in the most popular free tool for over a decade.
Modern schedulers, including the one Anki adopted in 2023 and the one Agent Memo uses, replace the hand-tuned formula with a model fitted to how people actually forget, so that a card's next review is placed where the chance of recalling it has fallen to a target the learner sets. The details are a subject of their own. What matters for using one is that it does two things: it grows the gap after a success and shrinks it after a miss, and it does this per card, so the thing you find hard comes back sooner than the thing you find easy.
Your part is the grade. After each attempt, you say how it went, and you say it honestly. A tool that lets you mark a card as known when you only recognised the answer will schedule it too far out, and the next time it appears you will have lost it. The four grades on most review screens, from "forgot" to "easy", are the only input the schedule has. Everything else is arithmetic.
Doing it by hand
You do not need software. A Leitner box with five compartments and a review rule runs a real spaced schedule and costs a pack of index cards.
| Box | Review every | A card moves here when |
|---|---|---|
| 1 | Day | It is new, or was just missed from any box |
| 2 | 3 days | It was recalled from box 1 |
| 3 | Week | It was recalled from box 2 |
| 4 | 2 weeks | It was recalled from box 3 |
| 5 | Month | It was recalled from box 4; retire it after two successes here |
Write a question on one side and the answer on the other. Each day, review box 1 and whichever other boxes are due. Attempt the answer before turning the card. A card recalled correctly moves up one box; a card missed goes back to box 1 regardless of where it was.
The system's weakness is the same as its strength: it is coarse. A card that was recalled only just and a card that was recalled instantly both move up one box. Software refines this by asking how the recall went and adjusting the gap accordingly. For a few dozen cards from a few books, the box is more than enough, and the physical act of moving a card is a surprisingly good motivator.
Doing it with software
Software earns its place when the number of cards passes what a box can hold, or when the cards live on your phone, or when you want the schedule to be finer than five compartments. As of September 2026, the choices split on three questions, and they matter more than the interface.
Who writes the cards. In Anki, Mochi and most classical tools, you do, one at a time. In Quizlet and RemNote, a generator will draft them from notes or a PDF. In Readwise, the cards are highlights you chose while reading. In Agent Memo, the agent reads the whole source and proposes the handful worth keeping, each traceable to its passage, and you keep or drop them before any enter review. The first is the most control and the most labour. The last is the tool we make, and the honest version of the trade is that it takes the card-writing off you and leaves the deciding.
What happens when you miss a week. This decides whether you are still using the tool in month three. Anki shows the whole backlog. Some tools cap the day. Agent Memo lets you set a ceiling of five, ten or twenty minutes and fits the schedule inside it, so a missed week comes back as a fuller session rather than a wall.
Whether you can leave. A tool that cannot export your cards and their history is a tool you cannot leave without starting over. Check for plain-text export and, if you are coming from Anki, for import that keeps the scheduling history rather than resetting every card to new.
A comparison of the current tools goes into each; the three questions are the ones to ask of any of them.
Where it breaks
The evidence for the method is solid. The evidence for people sticking with it is not, and the reasons are consistent enough to name.
The backlog. Reviews are due daily by design, so every day you skip adds to a pile, and the pile is what you see when you come back. A week away turns seven minutes into an hour, and an hour is a number that makes people close the app. This is the single most common way spaced repetition ends, and it is a design problem before it is a discipline problem. A tool that shows you a ceiling instead of a total removes most of it.
Too many cards. A method that works makes it tempting to put everything in it. Four hundred cards from five books means most of the daily session is spent on things you did not care about, and the cards you did care about are buried. The fix is upstream: fewer, better cards per source, and a habit of dropping a card the moment its answer stops mattering to you.
Cards written wrong. A card that can be answered by recognition, or that asks for a chapter, or that carries the date and not the reason, will be scheduled correctly and will teach you nothing. Scheduling is only as good as what it schedules. The active recall article is about writing cards that force the retrieval.
Week three. The first week is novel, the second is routine, and the third is when nothing feels different and the cards are from a book you have moved on from. There is no evidence the method is working yet, because the evidence is a thing you will still know in six months. The only protection is a session small enough that skipping it is not a relief, and a queue short enough that the cards in it are ones you wanted.
Spaced repetition for reading, not for exams
Most guides to this method are written for students, and the calibration for a student is not the calibration for an adult who reads. A student has a syllabus, a deadline, and a reason to keep three hundred cards for eight weeks. A reader has a shelf, no deadline, and a reason to keep a dozen things from each book for years.
For a reader, the schedule should be tuned for long retention, which means longer gaps and fewer cards. The daily session should be short enough to survive a job. The cards should be about ideas and mechanisms rather than facts, because a fact can be looked up and an idea cannot. And the card should carry where it came from, so that in a year you can check whether the number on it was the author's or yours.
Frequently asked questions
How long should the first gap be? About a day. The first drop is the steepest, so the first review comes soonest. After that, each success roughly doubles the gap in most schedulers, and a miss brings it back to a day.
Is spaced repetition the same as active recall? No. Active recall is producing the answer from memory; spaced repetition is the schedule for doing so. Spacing without recall, such as rereading a highlight every week, works far less well. Together they are the method.
How many cards a day is reasonable? For a reader, whatever fits in a session you will actually do. Twenty to forty reviews in seven to ten minutes is a common steady state once the schedule has matured. If the number keeps climbing, the problem is how many cards are going in, not the schedule.
Does it work for skills, or only for facts? It works for anything you can turn into a retrieval: a procedure, a distinction, the reasoning behind a decision, a word in a language. It does not replace the practice a physical skill needs. It keeps the knowledge the practice depends on.
Can I just review whenever I remember? You can, and it is better than nothing. But "whenever I remember" tends to mean either too soon, while the memory is fresh and the review is wasted, or too late, after it has gone. The schedule exists because people are bad at judging the gap, and the research says the gap is what matters.
What to do now
Take a dozen things you want to keep from the last thing you read, write each as a question, and either build the box above or put them into a tool with a daily ceiling. Review tomorrow, attempt before you look, grade honestly. If the part you know you will skip is writing the questions, let the agent propose them from the source, keep the ones you want, and spend the seven minutes on the reviewing instead.