The Forgetting Curve, Replicated

Most workplace training runs on an unstated assumption: what is taught in January is still there in June. The first person to test that assumption with numbers worked alone at a desk in Berlin, reading lists of invented syllables aloud to the beat of a metronome. What Hermann Ebbinghaus found in 1885 — and what a pair of Amsterdam researchers confirmed, curve for curve, 130 years later — is that the assumption fails in a specific and measurable way (Ebbinghaus, 1885).

This article walks through the original experiment, the 2015 replication that secured its place among the best-attested results in psychology, what the curve does and does not license you to conclude, and why a finding older than the automobile still describes the default outcome of most corporate training.

The experiment Ebbinghaus ran on himself

Ebbinghaus wanted to study memory uncontaminated by meaning, so he manufactured material that had none: thousands of consonant–vowel–consonant “nonsense syllables” — ZOF, DAX, WID — assembled into lists. He learned each list until he could recite it perfectly, waited a set interval, then measured what forgetting had cost him (Ebbinghaus, 1885).

His key instrument was the savings score. Rather than asking “how much can I still recall?”, he asked “how much faster can I relearn the list than I learned it the first time?” If a list originally took ten minutes and relearning took six, savings were 40%. Savings are a more sensitive measure than free recall, because they detect memory traces that have fallen below the threshold of conscious recollection.

Tested at seven intervals from about twenty minutes to thirty-one days, the numbers fell fast and then slowly: savings of roughly 58% after twenty minutes, about 44% after an hour, roughly a third after a day, and about 21% after a month (Ebbinghaus, 1885).

Two properties of that curve matter more than any single value on it. First, the loss is front-loaded: the steepest forgetting happens within hours, not weeks. Second, the curve flattens — what survives the first days is comparatively durable. A century later, an analysis that fit 105 candidate mathematical functions to 210 published retention datasets confirmed the same basic form across materials and methods: steep-then-flat, well described by logarithmic and power functions (Rubin & Wenzel, 1996).

A 130-year-old result, re-run

For most of the twentieth century the forgetting curve had a strange status: universally taught, almost never re-run in its original form. Ebbinghaus was his own only subject, his statistics were pre-modern, and his monograph predates the randomized experiment. When psychology’s replication debates intensified in the 2010s, the curve was an obvious candidate for a stress test.

Jaap Murre and Joeri Dros at the University of Amsterdam re-ran the 1885 study as faithfully as modern conditions allowed: a single participant, freshly constructed nonsense-syllable lists, the same savings measure, and the same seven retention intervals from twenty minutes to thirty-one days (Murre & Dros, 2015).

The replication succeeded. The shape of the 2015 curve tracks the 1885 curve closely — steep early loss, flattening tail — and the authors’ function-fitting confirmed that a simple exponential describes the data poorly, while power-family functions do well, consistent with the broader retention literature (Murre & Dros, 2015). The replication even reproduced a subtle anomaly: retention at the 24-hour point sits slightly above what a smooth curve predicts, a “jump” visible in both datasets that Murre and Dros suggest may reflect sleep-related consolidation (Murre & Dros, 2015).

Successful replications of 130-year-old experiments are vanishingly rare, and the authors were struck by how well the original held up — Ebbinghaus’s fixed daily schedule, controlled pacing, and obsessive procedural discipline anticipated methodological standards that would not be formalized for decades (Murre & Dros, 2015).

Beyond nonsense syllables

The standard objection is that nonsense syllables are deliberately meaningless, and meaningful material behaves better. The objection is half right. Meaning, structure, and prior knowledge slow the decay — but they do not repeal it.

The most striking evidence comes from Harry Bahrick’s studies of people who learned Spanish in school and were tested up to fifty years later: knowledge dropped steeply across the first few years after the last course, then stabilized into a plateau Bahrick called “permastore,” holding roughly level for decades (Bahrick, 1984). Same shape as Ebbinghaus’s curve — steep, then flat — stretched across a lifetime. In professional education the pattern repeats: reviews of knowledge retention in medical training find substantial loss of basic-science knowledge within the first year or two after study, even among high-performing learners (Custers, 2010). Meaning changes the slope of the curve, not its existence.

The most expensive ignored fact in training

Now hold the curve against the standard architecture of corporate learning: a workshop, a webinar, an e-learning module — one exposure, a completion checkbox, and no scheduled contact with the material ever again. The forgetting curve is a precise description of what happens to that investment next. It is not that one-shot training teaches nothing; it is that the design concedes the steepest part of the curve without a fight.

The irony is that the countermeasures are among the best-documented interventions in cognitive science. When Dunlosky and colleagues reviewed ten common learning techniques for efficacy, the ones that dominate practice — rereading, highlighting, summarizing — earned low-utility ratings, while the two rated high utility were exactly the two that attack the curve: distributed practice and practice testing (Dunlosky et al., 2013).

The evidence behind those two ratings is unusually deep. A meta-analysis of 839 assessments of distributed practice across 317 experiments found that spacing study over time reliably beats massing it in a single session (Cepeda et al., 2006). A follow-up study mapped the schedule itself, showing that the optimal gap between reviews scales with the retention horizon — the longer you need to remember something, the longer the ideal interval before revisiting it (Cepeda et al., 2008). And retrieval does more than measure memory; it strengthens it: students who practiced recalling a passage retained substantially more a week later than students who spent the same time restudying it (Roediger & Karpicke, 2006). The mechanism is what Bjork and Bjork call a desirable difficulty — conditions that make retrieval feel harder in the moment produce more durable learning than conditions that feel fluent (Bjork & Bjork, 2011).

None of this would matter operationally if the right review moment were easy to schedule by hand. It is not. The fade point differs by learner, by concept, and by how the last retrieval went — which is precisely why a 140-year-old descriptive law went unused for so long. Applying it requires bookkeeping at a per-person, per-concept granularity that no calendar, and no instructor, can sustain.

What the evidence doesn’t show

The forgetting curve is robust, but it is routinely over-claimed. Four limits are worth stating plainly:

• The viral numbers are not in the data. The claim that “people forget 70% of training within 24 hours” does not come from Ebbinghaus. His measure was savings on relearning, not percent forgotten, and his material was deliberately meaningless. His numbers do not transfer to meaningful, structured content — which decays more slowly, as Bahrick’s long plateaus show (Bahrick, 1984).
• It is an n-of-1 result, twice over. Both the 1885 original and the 2015 replication used a single subject (Murre & Dros, 2015). The generalization rests on the surrounding retention literature — hundreds of datasets showing the same functional form (Rubin & Wenzel, 1996) — not on these two studies alone.
• The curve is descriptive, not prescriptive. It tells you forgetting is lawful; it does not by itself tell you when to review. Optimal intervals come from the spacing literature and shift with the retention goal (Cepeda et al., 2008); there is no single schedule that is right for all material or all horizons.
• The rate is not a constant. Degree of initial learning, material type, sleep, and individual differences all change the slope. A single “company forgetting curve” applied to every employee and every topic is a convenient fiction; the law is in the shape, not in one universal decay rate.

From description to prescription

Ebbinghaus gave the field a law of loss; the replication confirmed that the law was real and not an artifact of one man’s heroic self-experimentation (Murre & Dros, 2015). The spacing and retrieval literatures then turned that description into a prescription: interrupt the curve with effortful retrieval, timed to the retention horizon (Dunlosky et al., 2013).

What remained missing for most of that history was the machinery to act on it — to track where each learner sits on each concept’s curve and intervene at the right moment. The science has been settled for longer than anyone reading this has been alive. The engineering, finally, is not the hard part anymore.