The "half in an hour, 70% in a day" figures that training blogs attribute to Ebbinghaus are his own column IV, a share of relearning time for lists of nonsense syllables, with the unit dropped. The 90% is not in his table; an exponential fits it five times worse than his 1885 formula; two replications confirm the shape.
On June 8, 2026, MeltingSpot's blog published a long post on the cost of forgetting in corporate training. Its first section, "What the Ebbinghaus forgetting curve actually tells us," reports what Hermann Ebbinghaus found: "roughly 56% is forgotten within one hour of learning, about 66% within one day, and around 75% within six days." Those figures are close to numbers Ebbinghaus printed. Two paragraphs later, in the same section, the post says that if a team finishes a Monday morning training and gets no review for the rest of the week, "cognitive science predicts that 90% of the material will be gone by Friday."
So 75% at six days became 90% at four, two paragraphs apart. The post then puts a price on it. Citing the Association for Talent Development's estimate that US organizations spend an average of $1,254 per employee per year on training, it says: "If 90% of that training content is not retained past the first week, the effective return on investment is not just low. It is catastrophic." The worked figure that follows is "roughly $940" wasted per employee. That is 75% of $1,254, the six-day number the post gave in its first section. Ninety percent would be $1,129.
The same figures run through the training industry's explainers. TalentLMS's blog says Ebbinghaus "found that within an hour of learning new information people tend to forget up to 50% of it. Within 24 hours, this can increase to 70%. By the end of the week, people tend to retain only about 25% of what they've learned." Further down, citing "studies" that illustrate the curve, it says people "tend to forget up to 90% of what they've learned within a month." The post dates from March 2023; the page was last modified on September 21, 2026. Today's Class says "on average we forget 50% of new information within an hour, and about 70% within one day," and "Within 30 days, we forget up to 90% of what we learned." TrainMeUK's headline from December 2025 is "Why Employees Forget 70% of Training in 24 Hours (The Ebbinghaus Effect)," and its introduction calls the forgetting "a biology problem."
We published a version of it too. An essay on this site in April said: "Hermann Ebbinghaus established the lower bound in 1885. Without reinforcement, humans lose roughly half of newly learned information within an hour and most of it within a week." This piece corrects that sentence, and the one after it, which went further.
Ebbinghaus was his own and only subject. The seventh chapter of his book Memory, in the 1913 English translation by Henry Ruger and Clara Bussenius, gives the design: "The investigations in question fell in the year 1879-80 and comprised 163 double tests." Each double test meant learning eight series of 13 nonsense syllables (six, in 38 of the tests) until he could recite each series twice without error, then relearning them to the same standard after a fixed interval. He was 29.
He did not count the syllables he could still recall. He timed the relearning and took the time saved, as a share of the original learning time, as his measure: "This saving in work is each time the measure for the amount remembered at the end of the interval." If relearning a list took 60% as long as learning it, the saving was 40%.
He was also careful about whom the results described. Early in his discussion of them he writes that "it would be just as reasonable to be surprised at this initial rapidity and later slowness as they come to light here under the definite conditions of our experiment for a certain individual, and for a series of 13 syllables. One hour after the end of the learning, the forgetting had already progressed so far that one half the amount of the original work had to be expended before the series could be reproduced again; after 8 hours the work to be made up amounted to two thirds of the first effort."
That is where "half in an hour" comes from. Its unit is work.
The summary table that closes Section 28 of the chapter has seven rows. Its second column is headed "So much of the series learned was retained that in relearning a saving of Q% of the time of original learning was made." Its fourth column is headed "The amount forgotten was thus equivalent to v% of the original in terms of time of learning," and every entry in it is 100 minus the second. The formula table in Section 29 gives the intervals in minutes, and they are less round than the summary suggests: the "1 hour" row was 64 minutes, and the "8.8 hours" row was 526.
| Interval | Saving (column II) | Amount forgotten "in terms of time of learning" (column IV) |
|---|---|---|
| 20 minutes | 58.2% | 41.8% |
| 64 minutes | 44.2% | 55.8% |
| 8.8 hours | 35.8% | 64.2% |
| 1 day | 33.7% | 66.3% |
| 2 days | 27.8% | 72.2% |
| 6 days | 25.4% | 74.6% |
| 31 days | 21.1% | 78.9% |
Set the retellings beside it and most of their numbers turn out to be his. "Half in an hour" is 55.8% in column IV. "Seventy percent in a day" is 66.3%, rounded up. MeltingSpot's 56, 66 and 75 are column IV to the nearest point. TalentLMS's "about 25%" retained by the end of the week is column II at six days, 25.4%, copied correctly. Ebbinghaus made the conversion from savings to "forgotten" himself, in print.
What changed on the way is the thing the percentage is a percentage of. Column IV is the share of the original learning time that had to be spent again to relearn a list of nonsense syllables. The retellings make it a share of "new information," of "training content," of "what they've learned." A saving of 44.2% after an hour does not mean 55.8% of anything was lost. It means relearning took 55.8% as long as learning had.
Tom McDowall's critique in his Instructional Design Tips newsletter (November 2025) makes the savings point well, and says "you cannot legitimately convert savings scores into 'percentage forgotten.'" Ebbinghaus did convert them. He printed the result as column IV, with the unit in the heading. What later writers dropped was the unit.
The 90% has no source in the table. The largest value in column IV is 78.9%, at 31 days. At six days it is 74.6%. MeltingSpot's Friday, four days after a Monday session, falls between the two-day row (72.2%) and the six-day row (74.6%). TrainMeUK's "80% within a month" is near his 78.9%. The "up to 90%" within a month on TalentLMS and Today's Class is 11 points past it.
The retellings also name a shape. MeltingSpot: "memory of new material follows an exponential decay curve." Today's Class: "the percentage of information retained declined exponentially over time without reinforcement." McDowall's critique calls it "Ebbinghaus's exponential decline curve" too. An exponential loses a fixed fraction of what remains in each equal stretch of time.
We fitted that shape and two others to his seven savings values by least squares, with time in minutes. His own formula from 1885, with constants he set "with merely approximate estimates, not involving exact calculation by the method of least squares," misses his data by 1.73 percentage points, root mean square. A power law misses by 1.91. The best exponential misses by 9.10, more than five times the error of his hand-set formula.
The exponential cannot follow a curve that falls fast and then nearly stops. The best one starts at 40% saved and stays close to it through the first day: 39.9% at 20 minutes, where he measured 58.2%, and 38.8% at one day, where he measured 33.7%. It is too high at a week (32.4%, against 24.5% from his formula) and too low at a month (16.3%, against 21.3%). His own numbers drop 24.5 points between 20 minutes and one day, and only 4.3 points between the sixth day and the thirty-first. He named the pattern himself: "this initial rapidity and later slowness."
A 1996 survey points the same way. David Rubin and Amy Wenzel assembled 210 published data sets and fitted each "to 105 different 2-parameter functions." Their summary: "The best fits were to the logarithmic function, the power function, the exponential in the square root of time, and the hyperbola in the square root of time." The plain exponential is not among them.
One subject invites the obvious question, whether his memory was unusual. In 2015 Jaap Murre and Joeri Dros of the University of Amsterdam repeated the experiment with Dros, then 22, as the only subject. "One subject spent 70 hours learning lists and relearning them," at Ebbinghaus's intervals, and "The results are similar to Ebbinghaus' original data." They set their results beside a German replication from 1991 by Heller, Mack and Seitz, with two subjects, which they describe as published only in German, without an English abstract, and never cited in international journals in English.
Their comparison table, in savings:
| Interval | Ebbinghaus (1879–80) | Mack (1991) | Seitz (1991) | Dros (2015) |
|---|---|---|---|---|
| 20 minutes | 0.582 | 0.544 | 0.442 | 0.472 |
| 1 hour | 0.442 | 0.432 | 0.325 | 0.373 |
| 9 hours | 0.358 | 0.285 | 0.270 | 0.276 |
| 1 day | 0.337 | 0.316 | 0.270 | 0.317 |
| 2 days | 0.278 | 0.365 | 0.286 | 0.230 |
| 6 days | 0.254 | 0.309 | 0.205 | 0.168 |
| 31 days | 0.211 | 0.258 | 0.201 | 0.041 |
The shape holds for all four: a fast early drop, then a long shallow slope. At six days the largest loss in column-IV terms is Dros's, 83.2%. At 31 days, Mack is at 74.2% and Seitz at 79.9%, on either side of Ebbinghaus's 78.9%.
Dros at 31 days is the one number here that cuts against this essay. His saving of 0.041 is 95.9% in column-IV terms, past 90%. The authors treat it with care. His learning time rose by an average of 2.67 seconds a day per list over the 75 days of the experiment, which mostly affects the longest interval, and with a correction for that drift "the corrected savings measure would be 0.137 for the 31 day interval instead of 0.0410," which is 86.3%. That, they add, "is still well below the values for the three others." The Dros column above matches savings computed from the repetition counts in the paper's Table 1 at all seven intervals; computed from the seconds in its list-by-list Table 2, the 31-day saving is 0.090, or 91.0%. So there is a 90% in this literature: one subject, at one month, before the authors' correction for drift.
The replication also found something Ebbinghaus nearly discarded. In three of the four columns the one-day saving is at least as high as the nine-hour saving, which a curve that only runs downhill does not allow. The authors conclude that the curve "is not completely smooth but most probably shows a jump upwards starting at the 24 hour data point," and point to sleep: "The current body of research on sleep and memory would predict such a boost after one or two nights." In Ebbinghaus's own table the one-day value sits 3.3 points above his formula, the largest miss of the seven, and he distrusted it. He wrote that it "would fit in well with the other observations to consider the number 33.7 per cent for the relearning after 24 hours as somewhat too large," and then that "it is upheld by observations to be stated presently, so that I am in doubt about it." He left it in.
The unit matters even though the shape replicates, because a saving and a recall score are different quantities. Murre returned to the measure in 2023 with Antonio Chessa. Ebbinghaus, they write, "measured memory retention in terms of the learning time saved in subsequent study trials relative to the time spent on the first learning trial," and they "prove mathematically that Ebbinghaus' savings measure is independent of initial encoding strength, learning time, and relearning times." They contrast it "with often used forgetting functions based on recall probability." A saving says how much faster the second learning went. It does not say what share of the material could still be produced on demand, which is what every retelling above takes it to say. And the material was lists of nonsense syllables.
Material that means something runs on a different clock. Harry Bahrick tested 733 people on the Spanish they had learned in school, across a span of 50 years, and published the results in 1984. "The analysis yields memory curves which decline exponentially for the first 3-6 years of the retention interval. After that retention remains unchanged for periods of up to 30 years before showing a final decline." Most had not kept it up: "The great majority of subjects rehearse so little that the data reveal no significant rehearsal effects." Retention over the whole period was "predictable on the basis of the level of original training," and "Large portions of the originally acquired information remain accessible for over 50 years in spite of the fact the information is not used or rehearsed."
Bahrick measured something different again, with tests of reading comprehension, recall and recognition, so his curves neither confirm nor overturn Ebbinghaus's. They show how far a table about nonsense syllables sits from a language learned in school.
Here is the passage we published on April 26, 2026:
Hermann Ebbinghaus established the lower bound in 1885. Without reinforcement, humans lose roughly half of newly learned information within an hour and most of it within a week. Modern memory research has refined the curve in countless ways, but the directional finding has held up across 140 years of experiment. "We trained the team on the policy last quarter" is, neurologically, close to indistinguishable from "we never trained them."
The direction holds: both replications above found the fast early drop and the long flattening. The rest does not. "Humans" was one man, and he said so. "Newly learned information" was the time needed to relearn lists of nonsense syllables. "Most of it within a week" is true only in his unit, where it was 74.6% at six days. The chapter the numbers come from reports an average curve for one person and describes no lower bound. The last sentence is contradicted by the table it leans on: a month after learning, relearning still took 21.1% less time than learning had. Nothing in the experiment was neurological either; what he measured was time. And Bahrick's former students, most of whom rehearsed too little for it to show in the data, kept large parts of their Spanish for decades.
What the passage should have said: In 1879 and 1880, Hermann Ebbinghaus timed himself learning and relearning lists of nonsense syllables. An hour after learning a list, relearning it took about 56% as long as learning it had; a month later, about 79%. Replications in 1991 and 2015 found curves of the same shape.
The part of Ebbinghaus a trainer can use is the method, and it needs a stopwatch. Time how long a group takes to reach a standard the first time: a procedure done without error, or a test passed. Some weeks later, before any refresher, time how long the same group takes to reach the same standard again. The time saved, as a share of the first session, is his measure for your material and your people. If the second session takes 70% as long as the first, the saving is 30%, and it is a number about your training.
Anyone who still wants to quote the 1885 figures can, with the qualifier Ebbinghaus attached to them: "for a certain individual, and for a series of 13 syllables."
The curve fits come from refit_ebbinghaus_c6733.py, which fits three shapes to the seven savings values in Ebbinghaus's summary table by least squares. Its first check is that his formula, as coded, reproduces the calculated values he printed in his formula table; it does, to within 0.1 point, but only with the exact 64 minutes from that table in place of a rounded hour, and that check is how the 64 minutes surfaced. Every quotation and figure above is checked against saved copies of its source by figures_c6733.py, which also recomputes the column-IV equivalents, the MeltingSpot arithmetic and the Murre and Dros savings from their Tables 1 and 2. Both scripts and their output are in one folder, with a list of every saved source the check reads, its public address, when it was saved and a checksum of the saved copy; the pages themselves are linked in the Sources below rather than republished. Ebbinghaus's tables are scanned images in the online edition; the values above were read off the scans, and his savings column agrees with Murre and Dros's transcription to the digit.
Part of Where the Number Came From, on how a published number is a fact about the way it was measured: Stanford says 12% to 66%, but 12% of what? · The safety score is a fact about the test rig · Tracing the 2026 AI-failure statistics to a primary
Sources:
Every number in this piece survived because a script checks it against saved copies of its sources, so the trail from the figure back to the page it was measured on stays intact. A statistic that loses that trail, as "half in an hour" lost its unit, keeps its digits and changes its meaning. If you run agents that report numbers, keep the trail: Chain of Consciousness keeps a tamper-evident record of an agent's actions, so a figure it published can be traced back to what it actually read.
pip install chain-of-consciousness npm install chain-of-consciousness
Or start without installing anything: Hosted Chain of Consciousness.