Ebbinghaus never said you forget 90% in a week

You've read the sentence. Some version of it opens every flashcard app's science page.

Within an hour you forget half. Within a day, 70%. Within a week, 90%. Ebbinghaus, 1885.

I read Chapter VII. Two of those numbers are close enough to defend. The third is off by a factor I checked twice.

Ebbinghaus published an equation. Solve it for 90% forgotten and you don't get a week.

Ebbinghaus's equation returns 5,309 years.

01 / 06

Both pages printing the 90% give no source for it

StudyCards AI, a flashcard company, on 3 September 2026:

StudyCards AI publishes a second page, "Exact Retention Percentages," which gets the measure right. Ebbinghaus "used a 'savings score': he measured how much less time it took to relearn the same list after a delay," with "33% remains accessible" at 24 hours.

Two pages, one site. The 90 appears only on the page that skips the definition.

TalentCards runs the same 50/70 opener and then swaps the window: "people tend to forget up to 90% of what they've learned within a month." Same 90, different deadline, no source given for either.

DigitalEd, the same day, prints "up to 50%" at an hour and "around 70%" at 24 hours. Then it says learners "often retain only about 25% of what they studied" by the end of the week.

That 25 is right, and it sits three sentences from the same borrowed opener.

And Wikipedia's forgetting curve article has been flagged since December 2018 as needing more citations. It carries the sentence that explains the whole mess: "'Savings' is thus, analogous to retention rate."

That single word, analogous, is where a measure of relearning time became a percentage of words you can say.

02 / 06

His own table never reaches 10%, at any interval

Ebbinghaus's full text is free. Chapter VII of the 1913 Ruger and Bussenius translation sits on York University's Classics archive.

The table lives in Section 28 as a scanned image. That's probably why so few people quoting the curve have looked at it.

Here's every row of it. Q is his savings percentage: how much of the original learning time he saved the second time. v is his own column for the rest, which he called the amount forgotten.

Read the 6-day row. 25.4% saved, 74.6% forgotten. The claim says 90 and 10.

You don't have to trust my reading of a scan, either. He summarises the same table in prose, one paragraph into Section 29:

One fourth at 6 days, in his own sentence. And look at what he says the half at one hour is: "one half the amount of the original work had to be expended."

Work. He's describing a bill for relearning, and it got copied down as a headcount of lost words.

Then read the last row. After a full month, Ebbinghaus still measured 21.1% savings.

His data never reaches 10% at any interval he tested, including one 5 times longer than the week the claim is about. So the 90 has no row to have come from.

Score the other two folklore numbers while we're here. He measured 55.8% forgotten at an hour against a claimed 50, and 66.3% at a day against a claimed 70.

Both are within about 6 points and in the right direction. The set is two-thirds honest, which is why the third goes unchecked.

03 / 06

Savings counts the seconds it takes to relearn a list

Ebbinghaus never tested how many syllables he could produce. He learned a list to two perfect recitations, waited, then learned it again, timing both.

His Section 28 definition is Q = 100D / (L - 85), where L is seconds of original learning and D is the difference.

Seconds. The unit is seconds.

Murre and Dros reran the whole experiment in 2015. They put the consequence about as bluntly as a journal will allow:

Zero percent recall and 100% savings, in the same person, on the same material.

Murre and Dros's concluding remark names what savings is closer to. Ebbinghaus's method "can also be seen as a precursor to implicit memory tests," picking up traces "which cannot be recalled or recognized but that do show savings." (An implicit memory is one that still speeds you up even when you can't bring it to mind.)

Cannot be recalled. The measure exists to detect the memories a flashcard scores as a failure.

The honest reading: after 6 days, Ebbinghaus needed about three quarters of his original study time to get a list of nonsense syllables back to perfect.

That's the finding. "You can say 25% of the words" isn't in it.

04 / 06

How long does Ebbinghaus's equation take to reach 90% forgotten?

About 5,309 years, if you push his own equation that far. Ebbinghaus fitted a curve to his seven points and published it in Section 29, with both constants:

I couldn't find anyone who had solved Ebbinghaus's equation for the number everyone quotes, so here's that arithmetic. Check it yourself.

The constants are his. The algebra is a bisection: guess a time, see which side of the target it lands on, halve the gap, and repeat.

First I checked that Ebbinghaus's equation reproduces his own printed column. He publishes calculated values of 57.0, 46.7, 34.5, 30.4, 28.1, 24.9 and 21.2.

Recomputing from the constants gives 57.0, 46.8, 34.5, 30.4, 28.1, 24.9, 21.2. The largest disagreement is 0.07 of a percentage point, which is his rounding.

Then I inverted it. Set b to 10, meaning 90% forgotten, and solve for t.

43 minutes, 1.1 days, then 5,309 years Solving his published equation for each of the three folklore thresholds. The first two land almost exactly where the story says; the third misses a 7-day week by a factor of about 277,000.

43 min

to reach 50% forgotten. The claim says an hour

1.1 days

to reach 70% forgotten. The claim says a day

5,309 yrs

to reach 90% forgotten. The claim says a week

Not measured data: this is arithmetic on a fitted curve. My own inversion of b = 100k/((log t)^c + k) with his published k = 1.84 and c = 1.25, t in minutes. 50% at t = 43, 70% at t = 1,614, 90% at t = 2,792,589,255. Ebbinghaus fitted this curve to 7 points spanning 19 minutes to 31 days, so the 5,309-year figure is extrapolation far outside his data and should be read as a statement about the equation, not about memory.

Two of the three solved figures land almost on top of the folklore. 43 minutes for the half, 1.1 days for the 70. Somebody, somewhere, did read the curve.

The third, 90% forgotten, is out by a factor of about 277,000 against a 7-day week. His function decays in the logarithm of time, so each step down takes far longer than the last.

Losing 90% of savings on a curve that flat takes a geological era. Anyone who plotted it would have watched the tail go horizontal and stopped.

The honest caveat, stated plainly because it cuts against my own headline: 5,309 years says nothing about human memory. It's what happens when you push a 7-point fit 60,000 times past its longest measurement. The man who wrote Section 29 would have refused to do that.

The point is narrower and sharper. On his own published curve, 90% forgotten and one week never meet, even with generous rounding.

05 / 06

Four subjects, a century apart, and the lowest is 16.8%

The obvious objection: one man in the 1880s, so who cares what his table says. Fair, and it has an answer.

Murre and Dros ran a full replication in 2015, open access in PLOS ONE: one subject, 70 hours on the same protocol. Their Table 3 puts four savings curves side by side. They're Ebbinghaus, Mack and Seitz from a German replication rarely cited in English, and their own subject, Dros.

At 6 days: 0.254, 0.309, 0.205, 0.168. The mean is 0.234.

The lowest of the four is 0.168, still 1.7 times the 0.10 the folklore claims. And that lowest value belongs to the subject Murre and Dros themselves flag as anomalous at the long end.

Four people, a century apart, in two languages. None of them lands anywhere near 10% at a week.

The replication turned up something the myth also flattens. Savings at 1 day and 2 days sit above the fitted curve in all four panels, by at least one standard error.

Murre and Dros conclude the curve "is not completely smooth but does show a jump at the 1 day retention interval." The sleep literature predicts that jump. But they're careful to say this experiment doesn't establish sleep as the cause.

A curve with a bump in it after your first night makes a worse slide and a better fact. It says the review you do tomorrow morning lands on a memory that just got stronger. We have a separate post on the sleep half of that.

06 / 06

Ebbinghaus refused to call his own curve a law

The last thing worth quoting is Ebbinghaus on his own results, since the pages citing him as a universal law never do.

And immediately after publishing the equation, in the same section:

Ebbinghaus calls his own curve a shorthand, and says he can't tell you whether the constants hold for anybody else.

Then he opens Section 30 with "even though only for my own case" and spends it checking two of his seven numbers against older tests. That's more caution in one chapter than the pages quoting him manage in a decade.

Ebbinghaus was learning nonsense syllables at 0.4 seconds a syllable, then timing how long it took to get back to perfect. You're learning Spanish words that mean things, hooked to sentences and a person who said them to you.

Whatever survives that trip survives as a shape. And a shape doesn't go in a headline as a percentage.

So what should you take? The curve is steep early and flat late, and that much has replicated four times. Reviewing sooner beats reviewing later.

That's the whole transferable finding. It's the same one retrieval practice and every modern scheduler already run on. They fit it to real review logs, where Ebbinghaus had seven points from one man's evenings.

Verbamor is one of the apps that cites this curve, so the correction is aimed at me too. What we do with it is narrow: cards built from your own lessons, scheduled by FSRS.

FSRS is an open-source scheduler. It learns your intervals from your actual answers, so it never assumes Ebbinghaus's constants apply to you.

If you keep your own Anki deck, it runs the same algorithm family for free, and a deck you maintain beats a subscription you abandon. I'd rather you use the numbers on this page to check my marketing than to trust it.

---

Sources