In someone else’s feed I came across a comment with numbers: Ebbinghaus’s curve, 1885—after twenty minutes, 58% of what was learned remains in memory; after thirty-one days, 21%. The figures looked credible, placed beside the name of a classic, and I was almost off to check where the gap between twenty minutes and a month had come from.
Something else caught me. In my archive of publications, there is not a single article on Ebbinghaus—despite its containing texts about how memory works and about how metrics outlive their authors. A subject everyone knows, never once examined. Which means it has a quality that lets it slip past: it looks so firmly established that there seems to be nothing to dig into.
It turned out that digging there means doing precisely what immortalized Ebbinghaus: making a habit of measuring what is inconvenient. And, incidentally, examining forty years of corporate training built on a number whose author appended the exact opposite caveat.
Hermann Ebbinghaus did not begin with memory. In 1873, he defended a philosophy dissertation in Bonn, worked independently for seven years, and throughout that time read Fechner’s Elements of Psychophysics. Fechner had measured how sensation depends on the intensity of a stimulus. Ebbinghaus decided the same trick might work with higher mental processes: if sensation has a law, why shouldn’t forgetting have one? He had Fechner’s book in English translation—historians still argue whether he bought it in Paris or in England.
He ran the experiments on himself. He chose his material slowly and fastidiously: tones proved too cumbersome, the digits from zero to nine too familiar, and stanzas from Byron’s Don Juan too variable in meaning. This was how nonsense syllables appeared: combinations like WID and ZOF, with neither meaning nor associations that might accidentally help or interfere. About two thousand three hundred such units in all; Roediger later clarified that the name was imprecise—there were real three-letter words among them too.
The main experiments took seven months, across 1879–1880, sometimes with three sessions a day. Ebbinghaus was 29. He was his own experimenter, subject, and statistician—and explicitly noted that he saw in this a source of error, which he called the “secret influence of theories and opinions,” and took precautions against it. In the forgetting experiment that concerns us now—163 “double tests.” Each one: eight series of thirteen syllables, learned to errorless recitation, then repeated after a preselected interval.
The intervals were these: roughly a third of an hour, one hour, nine hours, one day, two days, six days, and thirty-one days.
He was measuring not what people usually think. Not “how many syllables I remember,” but how much work relearning saves. The logic is simple: if a series took twenty-five repetitions to learn from scratch, and after a pause took twenty, then five repetitions were saved—20% savings. He called it “Ersparnis,” or savings in the English translation. His own formula, used to reduce the table: Q = 100 D / (L − 85), where 85 seconds for two control recitations were subtracted from the initial learning time.
The first reading disappoints anyone who loves a pretty graph. Ebbinghaus himself, in section 29, describes the result in words: after an hour, one had to spend again roughly half the original work; after eight hours, two-thirds. After that, the decline slows: “After 24 hours about one third was always remembered; after 6 days about one fourth; and after a whole month one fifth of the original work was completely saved.”
One third, one fourth, one fifth. 33.7% after a day, 25.4% after six days, 21.1% after a month—that is where “21%” comes from. Ebbinghaus himself gives the value for one day in the text, while discussing the oddity of that point.
And the numbers from that comment really are in his table. In reprints, the whole sequence circulates as 58.2 / 44.2 / 35.8 / 33.7 / 27.8 / 25.4 / 21.1. There is no error in them.
The error is elsewhere, and it is subtler: this is the share of effort saved, not the share of material retained. “21% after a month” means that relearning will take twenty-one percent less work than the first learning did—not “I remember twenty-one percent.” You cannot mechanically derive “what percentage was forgotten” from time savings: they are different quantities.
The shape of the curve is even more complicated. Ebbinghaus averaged the thirteen-syllable series as a whole, but the first two and last three syllables were learned easily—the primacy and recency effects showed through. The middle was hard going. A later replication confirmed it: the final curve is an average over elements of very different durability, some learned excellently and some only barely. And Ebbinghaus did not draw the graph of his famous curve: the book contains tables alone; the line is a later reconstruction.
Here is the sentence that ought to appear on every slide about the forgetting curve. Having finished fitting his formula, Ebbinghaus writes that it “has here no other significance than that of a condensed statement of the results given above, which have been found once and under the circumstances described.” Whether these results might carry a more general meaning, under other circumstances and with other people—he does not presume to say.
This is not a pose of false modesty. The experiment involved one person, and Ebbinghaus understood the price of a solitary observation. In 1985, for the book’s centennial, Roediger conducted an informal survey in his department’s hallways: everyone who had or was about to receive a psychology degree had heard of Ebbinghaus; a few could recount the idea of savings and the curve; and nobody—apart from a handful of “psychonomists”—had read the book. He put his conclusion bluntly: it is like describing Skinner’s contribution by saying, “He measured the responses of pigeons, fed them, and taught them to play ping-pong.”
The most interesting part of his experiments looked like a footnote. Ebbinghaus ran experiments not only with nonsense syllables but also with stanzas from Don Juan—and estimated that meaningful material required roughly one tenth of the effort needed for nonsense. In other words, the man whose name was given to the forgetting curve was the first to measure just how much the curve depends on meaningfulness, and found a difference of an order of magnitude.
And one more thing: he intended to continue researching memory, but never again did systematic work in the field, for unknown reasons. What followed were administration, journal editing, and textbooks. His 1909 obituary noted, not without venom, that “he was by no means inclined to hurry with his publications.”
A full reproduction of the experiment came only in 2015, when interest in reproducibility in psychology had become widespread. Jaap Murre and Joeri Dros of the University of Amsterdam repeated the procedure: one subject—Dros himself, age 22, a Dutch speaker—70 hours of work, 75 days instead of seven months, ten repetitions for each interval. The first non-German replication in one hundred and thirty years.
The result: the curve replicated. Its shape held, and that is the best thing one can say about scientific work of such age. With one exception—at thirty-one days, savings proved noticeably lower both for Ebbinghaus and for the two subjects in the 1991 German replication, which still has not been translated into English. The authors could not explain this and wrote plainly: perhaps they had fewer intervening lists between learning and relearning, and thus less interference, but there is nothing to confirm it—“this must remain speculation.”
But there was a surprise at one day. Ebbinghaus noticed it himself and filed it under things that cannot be: between nine and twenty-four hours, the decline was 2.5%, while between one day and two it was 6.1%, three times greater over a shorter stretch. He called this improbable and suggested that the 33.7% figure was simply overestimated by a couple of points, but hesitated because other observations supported that point. Separately, he mentioned night and sleep as a factor slowing the loss.
Murre and Dros: the one-day point does indeed rise above the prediction of the power function, and the most likely reason is memory consolidation during sleep. Most likely, not proven.
And the crucial thing the replication exposed: if you measure memory by the ordinary proportion of correct responses rather than effort savings, the curve becomes much flatter. After twenty minutes, the proportion of correct responses is just over 0.3 and barely falls after that. Two metrics for the same memory produce different shapes. Worse: the savings method cannot measure forgetting at all for material that is too easy. If a series is learned in one pass, then relearning it in one pass will yield one hundred percent savings after any interval—“complete absence of forgetting” will emerge as an artifact of the procedure.
That said, it is too early to bury the method. In 2022, Murre and Antonio Chessa showed mathematically that savings is a measure independent of the strength of the original learning and of time. In other words, a very “pure” measure. An old-fashioned, unwieldy instrument that spent forty years hammering nails that did not fit.
The corporate version of the figure has nothing whatsoever to do with Ebbinghaus, and its lineage has been lost for good. Its first documented stop is 1967: D. G. Treichler, an employee of Mobil Oil, publishes percentages such as “people remember 10% of what they read, 20% of what they hear, 30% of what they see” in the industry journal Film and Audio-Visual Communications—and cites not a single study.
A bare list of round numbers, paired up with perfect arithmetic: visual and auditory add up to the figure attributed to the two together. Even earlier, before the war, similar numbers surfaced at the Army’s Aberdeen ordnance school and in the oil industry alongside Paul John Phillips of the University of Texas. But that too is a trail, not a source. The NTL Institute in Bethel, Maine, still claims that he conducted the study and that the results were lost.
Then the numbers fused with Edgar Dale’s “Cone of Experience,” a 1946 diagram containing not a single figure and one which Dale himself warned should not be read literally. Molenda, Thalheimer, and coauthors published four articles in Educational Technology in 2014 dismantling this construction; criticism of it began back in 1971. In 2010, Thalheimer assembled 69 experimental conditions and more than a thousand subjects: people forgot anywhere from zero to ninety-four percent depending on context, and in the one-to-two-day window—from zero to seventy-three. His conclusion: any fixed-percentage “rules of thumb” are “simply false.”
Which did not stop the figure from taking on a life of its own. A 2011 training publication confidently reports: 42% is lost in twenty minutes and 67% in a day, cites Ebbinghaus, and dates his book to 1895. An industry blog in 2020 gives another pair: “60% of learning is forgotten after 20 minutes, 75% after 3 days.” The same curve, different numbers, not a single live citation.
A separate thread in the same story is the picture. Search for images of the “forgetting curve,” and most of what you find will be jagged sets of curves: each successive one flatter than the last because repetition compresses it. These are graphs from the SuperMemo spaced-repetition method, popularized by a 2008 Wired article. SuperMemo’s creator, Piotr Wozniak, wrote an analysis himself: he was not inspired by Ebbinghaus, and up to a third of the top search results bearing the German classic’s name have nothing to do with his data—including the image in his own encyclopedic article.
The difference is fundamental. Ebbinghaus studied decay without intervention and described it with a power function; SuperMemo curves show how decay changes under repetition. One squiggle versus another, with only the caption connecting them.