The forgetting curve: what Ebbinghaus actually found, and what to do about it
Somewhere in nearly every article about memory there is a graph: a line that plunges from the top left, flattens, and trails off toward the bottom right, usually with a percentage attached. Seventy per cent gone within a day. Ninety within a week. The graph is called the forgetting curve, it is attributed to Hermann Ebbinghaus, and it is used to sell everything from language apps to corporate training.
The curve is real. Most of the numbers written on it are not, or at least are not his. What Ebbinghaus measured was narrower and stranger than the infographic version, and what it actually tells you is more useful than the percentages, because it tells you when to do something rather than how bad things are.
This is what he did, what the shape means, what a modern replication found, which claims to distrust, and the single practical consequence that every serious review system is built on.
What Ebbinghaus did
Ebbinghaus published Über das Gedächtnis in 1885, and the experiments in it are among the first to treat memory as something that could be measured rather than described. He had one subject, himself, and one material: lists of nonsense syllables, consonant-vowel-consonant strings like DAX and ZUP, chosen because they carried no meaning and so could not be helped along by association.
He would learn a list until he could recite it twice without error, then leave it for an interval: twenty minutes, an hour, nine hours, a day, two days, six days, a month. Then he would relearn the same list to the same standard and record how much less time it took the second time. That saving, expressed as a fraction of the original effort, was his measure of how much had been retained.
Two things about this matter for reading the curve. The material was deliberately meaningless, so the results describe memory with nothing to hold on to. And the measure was savings on relearning, not recall: a list he could not recite at all might still be relearned faster than a new one, and that counted as retention.
The shape, and what it says
Plotted against time, his savings fell fast and then slowly. After twenty minutes he retained roughly three-fifths of the original effort; after a day, about a third; after a month, about a fifth. The drop from twenty minutes to a day was far steeper than the drop from a day to a month.
The percentages belong to nonsense syllables, one subject, and the savings measure. They do not transfer to the chapter you read last week, and anyone who tells you that you have lost a specific fraction of it by Thursday is quoting a shape as if it were a fact.
The shape does transfer, and the shape is the finding. Loss is front-loaded. Most of what will go, goes early; what survives the first few days has a much better chance of surviving the month. A memory is not draining at a steady rate. It is draining fast at first, and the rate itself slows as time passes.
That is the whole practical content of the curve, and it is enough to build a method on.
The useful response to early forgetting is to widen the gaps after successful recall.
What the 2015 replication found
For a study this widely cited, it was replicated remarkably rarely. In 2015, Jaap Murre and Joeri Dros repeated the original procedure: one subject, nonsense syllables, the savings method, the same intervals, with modern controls on timing and a great deal more care about the syllable lists. They published the results in PLOS ONE.
The shape held. Savings fell steeply in the first day and flattened after, and the fitted curve was close to Ebbinghaus's own. They also found something he had reported and later writers had mostly ignored: a small bump around the one-day mark, consistent with a night's sleep improving retention relative to the hours before it. And they noted, as he had, that the curve is not one curve but an average over syllables in different positions in the list, which forget at different rates.
The replication is worth knowing about for two reasons. It means the century-old result is not an artefact of one man's memory in 1880. And it makes the same caveat the original made: this is a curve for material with no meaning, learned to a criterion, measured by relearning. It is a model of the shape of forgetting, not a prediction for your reading.
The claims to distrust
Because the graph is simple and the name is famous, it attracts numbers it never contained.
"You forget X per cent within Y hours." These figures circulate with the graph and change from article to article, which is a sign that they are illustrative rather than measured. Ebbinghaus's savings after a day were around a third for nonsense syllables. What you retain from a paragraph with meaning, context and a reason for reading it is a different question, and it has no single answer.
"The curve applies to everything." It applies to isolated items learned to a criterion. Ideas embedded in an argument, skills you use, facts connected to things you already know: these are held up by more than the trace the syllables had, and they fall differently. The education researcher Carl Hendrick has argued that the curve is routinely stretched past what it shows, and the argument is fair. It describes a mechanism, not a syllabus.
"Reviewing at the right times makes the curve go flat." Partly true, and the part that is true is the useful part; see below. But the flattening is per item and per review, and it is a slowing of the decline, not a stop. Nothing in the data says a memory becomes permanent.
"Ebbinghaus discovered spaced repetition." He noticed the spacing effect, which is that repetitions spread over time were worth more than the same repetitions massed together. He did not build a schedule from it. That came almost a century later, with the Leitner box in the 1970s and SuperMemo's algorithms in the late 1980s.
The one thing the shape tells you to do
If loss is front-loaded, then the moment to act is early, and the useful action is a review timed to land just before the memory would have gone.
Here is the mechanism, in words rather than a diagram. Learn something today; by tomorrow, without a review, a large part of it is at risk. Review it tomorrow, recalling it rather than rereading it, and two things happen. The memory is restored, and its curve going forward is shallower than the one it had yesterday: the next drop is slower. Review it again a few days later, at the point where it has begun to fade, and the curve after that is shallower still. Each well-timed review buys a longer gap before the next one is needed.
The consequence is a schedule that grows: a day, then a few days, then a couple of weeks, then months. That is spaced repetition, and it is the forgetting curve read as an instruction. The review has to be a recall attempt, because a reread does not restore a memory in the same way; that is the subject of the active recall article. And the timing does not have to be perfect. A review that comes a little early is slightly wasted; one that comes a little late is a miss that resets the gap. Both are recoverable. What is not recoverable is never reviewing, which is what reading normally involves.
A rough picture of what the curve looks like for a single item under three regimes, described in words because the percentages would be invented:
| Regime | Day 1 | Week 1 | Month 1 | Month 6 |
|---|---|---|---|---|
| Read once, never reviewed | Most of it at risk | Most of it gone | The gist, if that | Recognition at best |
| Reread once, a week later | Same as above | Briefly restored | Most of it gone again | Recognition |
| Recalled at 1 day, 4 days, 2 weeks, 6 weeks | Restored | Restored, curve shallower | Restored, gap now months | Likely still recallable |
The third row is the only one where month six looks different from month one. It costs four short sessions, each a minute or two per item, timed by the shape of the curve rather than by whenever you happen to think of it.
For a reader, not a subject
If what you want is to keep what you read rather than to reproduce a syllable list, three adjustments to the picture.
Meaning slows the curve before you do anything. A passage you understood, connected to something you already knew, and read because you had a question about it starts from a shallower slope than DAX ZUP. The steep early loss is still there; it is just less total.
You do not want everything to survive. The curve describes each item on its own. A book has a dozen items you would want in a year and several hundred you would not, and the whole method fails if you try to flatten the curve for all of them. Choosing the dozen is the work that precedes any schedule; there is a method for it.
And the first review is the one you are least likely to do, because it comes tomorrow, when the book is still fresh and the review feels pointless. It is not pointless. Tomorrow is where the curve is steepest, and the review that lands there is the one that buys the most time.
Frequently asked questions
Did Ebbinghaus really find that we forget 70 per cent in a day? No. His measure was savings on relearning nonsense syllables, and after a day the saving was around a third of the original effort, which is not the same as forgetting two-thirds. The specific percentages attached to the graph in most articles are illustrations, not findings.
Has the forgetting curve been replicated? Yes. Murre and Dros repeated the original procedure in 2015 with one subject and modern controls, and found the same shape, including a small improvement around the one-day mark consistent with sleep.
Does the curve apply to things I understand, or only to nonsense? The shape applies broadly; the numbers do not. Meaningful, connected material starts from a shallower slope and loses less in total, but the loss is still front-loaded, which is the part that matters for timing reviews.
Can a memory be made permanent? Nothing in the data says so. Well-timed recalls make the curve shallower each time, until the gap between reviews is measured in months or years. That is as close to permanent as the method gets, and for most purposes it is close enough.
What is the best time for the first review? About a day after learning. That is where the curve is steepest and where a recall attempt buys the most. After that, each successful review earns a longer gap.
What to do now
The next time you read something worth keeping, do not draw the graph; act on it. Write a question from it, and tomorrow, before you look, try to answer. Then let a schedule with a daily ceiling decide when it comes back, or have the questions proposed from the source and keep the ones you want. The curve does the rest.