The Half-Life of Knowing

The follow-up question is the practical one: how far apart? The answer is the elegant part. The best gap is not fixed; it scales with how long you need to remember. In a large study that varied both the gap between two sessions and the delay until the final test, the optimal inter-study gap came out as a sizeable fraction of the retention interval — on the order of 10 to 40% of it at week-scale delays, shrinking toward roughly 5% when the target was a year away 4. To remember something for a week, review it after a day; to remember it for a year, review it after a month or two. The schedule should breathe in proportion to the horizon. This is why a good spacing system is not a single magic interval but a widening one — a point the algorithms will make concrete.

There is a second engine of durable memory that rides alongside spacing, and the two reinforce each other: the testing effect. Trying to retrieve a fact — as opposed to re-reading it — is itself one of the most powerful study acts known. Taking a test on material, even with no feedback, produces more durable retention than restudying it for the same time 7, an effect so strong that its authors called retrieval not merely a way of measuring learning but a primary cause of it 8. Spaced repetition is, at heart, a machine for scheduling retrievals. It does not ask you to re-read. It asks you to recall — and it picks the moment.

Why a little forgetting helps

So why should spacing the retrievals matter, beyond simply doing more of them? The deepest answer is a productive paradox. A memory retrieved easily — just after studying, while it is still fresh — is barely strengthened by the act. A memory retrieved with effort, just as it begins to slip, is strengthened a great deal. The struggle is not a tax on learning; it is the signal that drives it.

Robert and Elizabeth Bjork gave this a formal frame in their new theory of disuse 5, which splits a memory into two quantities that are easy to confuse and important to separate. Retrieval strength is how accessible an item is right now — high just after study, fading with time. Storage strength is how deeply learned it is — how well it will come back after a delay. The theory’s sharp, counter-intuitive claim is that the gain in storage strength from a successful retrieval is larger when retrieval strength has fallen lower. In other words: the more you have begun to forget — provided you can still, with effort, recall — the more a successful recall is worth. Robert Bjork named the broader principle a desirable difficulty 6: conditions that make practice feel harder and slower can make the resulting learning more durable.

The paradox has a hard limit, and the figure above is built around it. Push the difficulty too far — wait until the memory is genuinely gone — and the retrieval simply fails. Now you are not strengthening anything; you are relearning from scratch, which is slow and dispiriting. The probability that you succeed at the recall is, after all, just the retrievability itself. So two pressures pull against each other: the boost from a successful review grows as the memory fades, but the chance of a successful review shrinks at the same time. The sweet spot is an intermediate retrievability — fade the memory enough that recalling it is real work, but not so much that the work fails. Cognitive models of the spacing effect, such as Pavlik and Anderson’s activation-based account built on the ACT-R architecture, reproduce exactly this trade-off from first principles, and predict the spacing benefit as a function of the gap 9. The whole art of a scheduler is to keep landing reviews in that window.

Turning it into a schedule

Once you believe all this, an algorithm almost writes itself: track each item’s memory, predict when its retrievability will fall to the edge of the useful window, and schedule the next review for that day. The first systems to do it were beautifully analogue. Sebastian Leitner’s 1972 box method needs nothing but index cards and a row of boxes 12: a card you answer correctly moves to the next box along, which is reviewed less often; a card you miss falls back to the first. The boxes are the stability estimate, made of cardboard.

The computational leap came from Piotr Woźniak, whose SuperMemo and its SM-2 algorithm, devised in the late 1980s, still underpins much of the field 13. SM-2 keeps, for every item, an ease factor — a per-item multiplier that starts at 2.5. The first review interval is one day, the second is six, and every interval after that is the previous one times the ease factor:

$I_n = I_{n-1}\times \mathrm{EF}, \qquad \mathrm{EF} = 2.5.$

Each time you grade your own recall, the ease factor is nudged — down for a hard item, up for an easy one, never below a floor of 1.3 — and a lapse sends the item back to the start. The consequence is an expanding schedule, exactly the widening rhythm the spacing research called for: one day, six, fifteen, thirty-eight, ninety-five, and on.

Notice what the expansion buys. An item you keep getting right demands your attention less and less often; the schedule thins itself out, freeing time for the things you are still struggling with. Woźniak and Gorzelańczyk later set the approach on firmer theoretical ground, framing the spacing of repetitions as an explicit optimisation of long-term retention against review effort 14. SM-2 was a hand-built rule of thumb that happened to capture the shape of memory. The next step was to stop guessing the shape and learn it.

A memory you can fit to data

A hand-tuned ease factor is a strong guess about how memory works. But a flashcard app logs millions of reviews — every card, every grade, every interval, every success and lapse — and that is precisely the raw material a statistical model wants. The modern turn in spaced repetition is to fit the forgetting curve itself to data.

The most widely used open implementation is FSRS, the Free Spaced Repetition Scheduler, now built into the flashcard program Anki as of its 23.10 release. It models every item with three numbers, and the diagram below is their whole story. Difficulty is how stubborn the item is to make stick. Stability is how long the memory lasts — defined, cleanly, as the interval at which retrievability falls to exactly 0.9. Retrievability is the live probability of recall, computed from the elapsed time and the current stability. After each review the model updates the trio: a recalled review increases stability — by more when retrievability had dropped low, the desirable-difficulty principle made quantitative — and a lapse cuts stability and raises difficulty 18. The parameters that govern those updates are not guessed; they are fit by gradient descent to a learner’s own review history.

The same instinct — learn the memory model from logs — has been pursued at industrial scale. Duolingo’s half-life regression trains a model to predict the half-life of each word for each learner, fit on the order of thirteen million review logs, and uses it to schedule practice 15. And the payoff is not confined to apps: in a semester-long classroom study, eighth-graders learning Spanish vocabulary on a personalised review schedule — one that modelled each student’s memory and timed reviews accordingly — outscored a massed schedule by 16.5% and even a generic spaced schedule by 10% on a cumulative exam a month after the semester ended, a gap of roughly two letter grades on the early-term material 16. Scheduling, done well, is not a marginal optimisation. It moves grades.

At the research frontier the problem has been posed in its most general form, as one of optimal control: given a model of how reviews move a memory, and a cost on the learner’s time, what review schedule maximises long-term recall? Casting it as stochastic optimal control yields scheduling policies with provable guarantees, and — pleasingly — they often recommend reviewing at a roughly constant target retrievability, the very window the desirable-difficulty argument pointed to 17. The hand-built rule and the optimal policy converge on the same advice.

The shape of forgetting

We have been writing the forgetting curve as an exponential, and it is time to admit that this is probably wrong — interestingly wrong. Ebbinghaus’s curve is commonly drawn as an exponential, but a long line of work argues that long-term forgetting follows a power law instead, with a much heavier tail. Wickelgren proposed power-function forgetting in the 1970s, and Wixted and Carpenter later showed that a power law of the form $R = m(1 + ht)^{-f}$ fits the savings data — Ebbinghaus’s own included — better than a simple exponential 1011.

The difference is not academic hair-splitting; it changes the schedule. An exponential decays at a constant proportional rate forever, which means a memory you have not reviewed in a year is essentially gone. A power law decelerates: the older a memory gets, the more slowly it fades — an observation old enough to have a name, Jost’s law. Anchor the two curves at the same stability and they agree in the short run and diverge sharply in the long run, the power law holding a recallable trace long after the exponential has flatlined. This is why the serious modern schedulers — FSRS among them — fit a power forgetting curve rather than the textbook exponential: its shape is closer to how human memory actually behaves, and it forgives the long gaps that real revision schedules inevitably contain.

Where the model thumbs the scale

It would be easy to read all this as a solved problem, and it is not. The honest reading is that scheduling memory is a genuinely useful tool with genuine limits, and the limits matter.

First, every model here predicts a population, not your particular brain on this particular Tuesday. Stability and difficulty are fitted averages; an individual recall is a coin weighted by them, not determined by them. A scheduler that has seen only a handful of your reviews is mostly guessing — the cold-start problem — and grows trustworthy only with history.

Second, and more fundamentally: a scheduler decides when to bring a fact back, but it cannot do the bringing-back for you. The entire benefit rests on the assumption that the review is an effortful retrieval 78. Flip a card, sigh, and read the back without trying, and the desirable difficulty collapses into no difficulty at all; the interval expands on the strength of a recall that never happened. The algorithm is only ever as good as the honesty of the act it schedules.

Third, spacing optimises retention of specified items, which is not the same as understanding. It will faithfully keep a definition, a date, a formula at your fingertips. It will not, by itself, teach you to connect them, to transfer a method to an unfamiliar problem, or to see why a result is true. Other desirable difficulties — interleaving different problem types rather than blocking them, which improves the ability to choose the right method 19 — address parts of what pure spacing leaves out. A revision strategy that is all flashcards and no synthesis has optimised one real thing while quietly neglecting another.

And finally there is a human limit the mathematics is silent on. An aggressive scheduler can pile up a punishing daily backlog of due reviews, and a learner buried under it abandons the system entirely — at which point the theoretically optimal schedule retains nothing, because no one is following it. The best schedule a person actually keeps beats the perfect one they quit.

What we read from it

EuraStudy is built on a reviewed bank of exam-style questions, each tagged to a topic and a curriculum — and a question, retrieved at the right moment, is exactly the effortful recall this whole literature turns on. The spaced-repetition lens gives us a way to think about that bank not as a static archive but as a set of memories, each with its own quiet half-life, each due to be brought back before it fades.

The reading we take from a century of this work is threefold, and bracingly simple. Forgetting is not failure — it is the gradient the whole method climbs; a review that costs you a little effort is worth more than one that costs you none. When beats how much — the same practice, timed to the edge of forgetting, outlasts far more practice crammed together. And the schedule should widen — an item you keep recalling has earned a longer leash, and the time saved belongs to the things you have not yet learned.

There is a tidy symmetry in where this leaves the series. Item response theory gave us the snapshot — how able you are, at one instant. Knowledge tracing gave us the film — how that ability moves as you learn. Spaced repetition gives us the calendar — when the film needs replaying before a scene is lost. None of the three does the learning for you. Between them, they decide what to ask, judge how you answer, and choose when to ask again — and then leave the one irreducible act, the act of remembering, where it has always belonged.