You learn something. A word, a fact, a name. For a day or two it feels solid. Then, without any single moment you could point to, it's gone - not wrong, just unavailable, like it was never filed anywhere you can reach.
That's not a memory failure specific to you. It's the default behavior of memory itself, and it's been measured precisely for well over a century.
In short
- Memory fades quickly after learning, then more slowly - and it keeps fading unless something interrupts it.
- Cramming helps with tomorrow's test and does almost nothing for next month's recall.
- Spaced repetition means reviewing each item at the moment you are about to forget it, so each review pushes the next one further out.
- Reviews work best when you try to pull the answer out of memory before you see it.
- Better schedulers time reviews per item, using a model of your memory instead of one table for everyone.
The forgetting curve
In the 1880s, the psychologist Hermann Ebbinghaus ran a simple, brutal experiment on himself: memorize lists of meaningless syllables, then test his own recall at increasing intervals afterward. The result, plotted out, is now called the forgetting curve - and the shape is steep. A large chunk of what you learn is gone within a day. What survives that first drop decays more slowly, but it keeps decaying, indefinitely, unless something interrupts it. A 2015 replication using the same method found the same basic curve, which is a strong sign this is a property of memory, not of one man's notebook. The full story is in the forgetting curve.
The uncomfortable part isn't that we forget. It's how fast the first forgetting happens - faster than most people's intuition expects, and far faster than "I'll remember this" accounts for.
A concrete example
Say you meet a new word in a language you're learning - Schlüssel, key. On day one you see it in a lesson, use it twice, and feel confident. Two days later someone points at their keys and you reach for the word and find nothing. You didn't fail to learn it. You learned it, and the curve did what the curve does.
Now imagine a second version of the same week. You see the word on day one. You're asked for it again on day two, right when it's getting hard to retrieve, and you manage it with a little effort. You're asked again four days later, then ten days later. By the third check, the word is stable enough that you could go a month without seeing it. Same word, same learner. The only difference is when the reviews happened.
Why cramming doesn't fix it
Cramming - reviewing something intensively right before you need it - works for exactly one thing: the test tomorrow. It does nothing for the curve itself. You've just moved your one data point earlier; the material still decays at the same rate the moment you stop looking at it, and a week later it's gone the same way it would have been anyway.
Cramming doesn't flatten the forgetting curve. It just moves your test date to before the curve drops.
The fix isn't reviewing more. It's reviewing at the right moments - specifically, close to the point where you're about to forget, not right after you just learned it and not so late that you've already lost it.
This is one of the oldest results in the psychology of learning. Ebbinghaus is credited with first identifying the spacing effect, and later work pooling many experiments found spaced practice beating massed practice in the great majority of comparisons (overview of the spacing effect).
Why spacing it out works
This is where spaced repetition comes in, and the mechanism is almost too simple to sound credible: each time you successfully recall something right as it's about to slip, the next forgetting curve for that item is flatter. It decays slower. Review it again at its new, later at-risk point, and the curve flattens further. Do this enough times and an item that used to need daily review might only need to be seen once a month to stay solid.
A useful way to picture it is a leaky bucket you top up. Fill it right after learning and it drains fast, so the first top-up has to come soon. Each top-up that works makes the bucket leak a little more slowly, so the gaps can stretch: a day, then several days, then weeks, then months. The total number of reviews over a year ends up small compared with how much you keep.
The catch is that "right as it's about to slip" is a moving target - it's different for every item, and different for every person, because it depends on how well you actually know that specific thing right now. A word you've seen fifty times and a word you've seen twice don't decay at the same rate, even if you learned them on the same day.
What a review actually is
Timing is half the story. The other half is what you do during the review. Looking at the answer and nodding along is a weak form of review. The version that works is trying to retrieve the answer from memory first, then checking. That effort is part of what makes the memory stronger, and it's usually called active recall or the testing effect (overview of the testing effect).
Flashcards are popular for a reason: they force the attempt. Cover the back, try to produce it, then flip. If you got it, the next gap grows; if you didn't, you see the answer and the card comes back sooner. Active recall vs spaced repetition goes into why these are two different ideas that work best together.
Where the schedule comes from
For a long time, the practical answer to the "moving target" problem was a fixed table: review a new item after a day, then three days, then a week, then two weeks, the same sequence for every card regardless of how you actually did on it. Some of these systems are simple enough to run with a shoebox, as in the Leitner system. Others adjust intervals with an "ease" number per card, as in the older algorithm compared in FSRS vs SM-2. All of them are reasonable approximations, and all of them beat having no schedule at all.
What changed with better scheduling
What's changed more recently is scheduling that models the curve directly, per item, updated by how you actually perform - not a fixed table applied uniformly, but an estimate of exactly when a specific piece of information, for you specifically, is about to be forgotten. What is FSRS explains the model. Cirello's scheduling is built on FSRS-6: reviews aren't timed by a generic interval, they're timed by an ongoing model of your own forgetting curve for that exact card, with parameters retrained nightly on your own review history (inside the pipeline).
Making it a habit instead of a project
Knowing how it works is not the hard part. The hard part is the two things that make people stop: the effort of making cards and the guilt of a growing backlog. Both are design problems, not discipline problems.
- Keep capture cheap. If a list already exists, paste it in rather than rebuilding it card by card.
- Keep cards small. One fact per card is faster to answer and easier to schedule - see how to write good flashcards.
- Keep the daily load honest. A small number of new cards each day produces a manageable queue; a huge batch produces a wave of reviews weeks later. How many flashcards per day shows how to think about it.
- Make missing a day cheap. A good tool lets you pick up where you are instead of making you pay back a debt.
If you want to see how this plays out for a specific kind of learning, there are pages for vocabulary and language learning, dense professional material, and remembering what you read. The general overview of the tool is on the spaced repetition flashcard app page.
What spaced repetition does not do
It's worth being plain about the limits. Spaced repetition doesn't teach you to understand something - it keeps what you've already understood from evaporating. Cards built on material you don't understand tend to become rote strings that are hard to apply. It also doesn't make remembering effortless; reviews should feel like mild, honest effort, not like checking a box.
And it isn't a reason to put everything in. The best decks are selective: things worth carrying around in your head for months or years.
The forgetting curve isn't a flaw to work around. It's just how memory behaves - and once you can see its actual shape, timing review around it stops being guesswork.