The Forgetting Curve: One Man, Nonsense Syllables, and the Percentages Nobody Can Trace

You will have seen the graph. A line falls off a cliff in the first hours and then flattens, with percentages written beside it: seventy per cent gone in a day, ninety per cent within a week. It is the standard opening slide for spaced-repetition apps and corporate training platforms, and it is always attributed to Hermann Ebbinghaus in 1885.

The curve is real. It has replicated. And almost everything attached to it in that slide is not what Ebbinghaus measured.

What the original experiment was

Ebbinghaus (1885) had one subject. It was himself.

The details come from Murre and Dros (2015), in PLOS ONE 10(7), e0120644, who reconstructed the procedure in order to repeat it. Ebbinghaus learned lists of nonsense syllables - meaningless three-letter strings, arranged in rows of thirteen - then relearned them after set intervals and recorded how much less effort relearning took. That measure is called savings, and it is not the same as recall: it is how much of the original learning work you are spared the second time.

The retention intervals were roughly 20 minutes, 1 hour, 9 hours, 1 day, 2 days, 6 days and 31 days. The work took seven months, often with up to three sessions a day.

So the founding data set is one man, memorising meaningless syllables, measured by relearning effort. Every property of that design matters for what the curve can be used to claim.

It replicated - with one subject again

Murre and Dros did something unusual and admirable: they ran it again properly.

Their subject was Joeri Dros, aged 22, the study’s second author. He learned 70 lists of 104 nonsense syllables each (eight rows of thirteen) over 75 days, from December 2011 to February 2012, investing roughly 70 hours of learning time, with ten replications at each retention interval (nine at the nine-hour condition).

The result held. Comparing their curve with Ebbinghaus’s and with two earlier replications, the authors write: “we find the resemblance of the four graphs remarkable.”

There is one deviation, and they report it rather than smoothing it. At the 31-day interval, Dros’s savings score was 0.041, against roughly 0.20 in the other data sets. The authors do not claim to know why, offering “pro-active interference or fatigue” as possibilities.

And there is a structural finding that contradicts the picture on the slide. Their conclusion: “the Ebbinghaus forgetting curve has indeed been replicated and that it is not completely smooth but most probably shows a jump upwards starting at the 24 hour data point.”

The curve everyone draws is a smooth exponential decay. The best replication of it says there is a bump at twenty-four hours - probably a sleep effect - and that the line is not smooth.

There is no single curve of forgetting

The deeper problem with the slide is that it implies one law.

Rubin and Wenzel (1996) tested that directly. They assembled 210 published retention data sets and fitted 105 different two-parameter mathematical functions to each, requiring at least five data points and a fit correlating at .90 or better.

Four functions fitted comparably well: logarithmic, power, exponential in the square root of time, and hyperbola in the square root of time. The authors conclude that “it is difficult to distinguish among these 4 functions”, and that the same four fitted most data sets - with autobiographical memory as the exception.

That is not a single universal forgetting curve. It is a family of functions that describe forgetting about equally well for most kinds of material, and at least one kind of memory that behaves differently.

Where the percentages come from

Search this term and you will meet a genre. Corporate training and e-learning vendors state, as established fact and with Ebbinghaus’s name attached, that employees forget seventy per cent of training within twenty-four hours and ninety per cent within a week.

We went looking for the study that establishes those figures for training content, and we did not find one.

What exists is a savings curve for nonsense syllables, produced by one person in the 1880s and replicated by one person in 2011. Savings on meaningless material is not the same quantity as recall of meaningful material, and a single subject’s curve is not a population estimate. Transposing a number off that graph and onto “what your employees retain from a compliance course” is a category change, not a citation.

We are therefore not printing a percentage for how much Ebbinghaus retained at twenty-four hours, even though it would be easy to. We could not verify it in a source we could read, and the number would immediately be used for exactly the transposition this section is about.

What survives, and what it is good for

Quite a lot survives, and it is more useful than the slide.

Forgetting is steep early and slows down. That is the shape, it replicated across a century, and it is the reason reviewing soon after learning is better placed than reviewing much later. There is probably a discontinuity around the twenty-four-hour mark. And forgetting is well described by a small family of curves rather than one.

What does not survive is the specificity. There is no verified percentage for what you will retain from a lecture tomorrow, because the curve was never a measurement of that.

This is a claim about the curve itself. Whether spacing your practice actually improves retention, and whether the scheduling rules attached to it hold up, is a separate question with a much larger evidence base and its own set of overstated rules - and that article is where the practical question belongs. The reason rereading feels productive and does not work is covered in the guide to studying.

Limitations and disputes

• Ebbinghaus (1885) was not read in the original. Everything here about its design comes from Murre and Dros (2015), who reconstructed it in order to replicate it.
• Both the original and the replication have one subject. The replication is a genuine achievement and it is still n = 1. The authors are explicit about this being a single-subject design.
• The 31-day anomaly is unexplained by the authors themselves.
• We have not printed any retention percentage attributed to Ebbinghaus, for the reasons given above.
• Rubin and Wenzel’s journal details could not be fully confirmed; volume and pages are omitted rather than guessed.
• Nonsense syllables are the point and the problem. Ebbinghaus chose them deliberately to strip out prior knowledge. That is what makes the curve clean and what makes it a poor model of learning things that mean something.
• This article does not evaluate any spaced-repetition product, and none of the above supports or undermines a specific app.