The Curve of Forgetting

A scientist named Hermann Ebbinghaus wanted to know how fast people forget. He made up nonsense words like ZUV and KEB — words with no meaning to help him remember — and learned long lists of them by heart.

Then he waited — an hour, a day, a week — and tested himself. He forgot nearly half within the first day, but after that the forgetting almost stopped. Fast at first, then slow: that shape is called the forgetting curve. Practicing a list again made its curve fall more slowly the next time.

Ebbinghaus needed material nobody could already know, so every list started equally new — that is why he invented nonsense syllables. And he needed a measure finer than remembered-or-forgotten, so he invented the savings method: relearn the list to the same standard and compare the effort. If learning took 20 minutes and relearning took 13, then 7 minutes were saved — 35 percent savings. Savings can detect memory that recall misses entirely.

His results: about 58 percent savings after 20 minutes, 44 after an hour, 34 after a day — then the curve nearly levels, still holding 21 percent a month later. Forgetting is steep in the first hours and patient afterwards. Each well-timed review resets the curve and, crucially, flattens it: after every relearning, the same delay costs less.

Three design choices made the experiment work. Nonsense syllables stripped away prior knowledge, so every item started equal. Learning to a fixed criterion — two perfect recitations — made lists comparable. And savings, computed as learning time minus relearning time, divided by learning time, gave a continuous measure sensitive enough to register memory that felt completely gone: a list he could no longer recall at all still relearned faster than a fresh one.

The curve that emerged falls roughly like a logarithm — the loss per unit of time shrinks as time passes — and it replicates: in 2015, Murre and Dros repeated the full protocol on a modern subject and recovered Ebbinghaus's shape almost point for point. The practical lever is the reset: a review timed to land before the curve bottoms out restores full strength and leaves a shallower slope behind it.

You learn a new song today. Tomorrow, some of it is gone. Forgetting happens fastest right after you learn. Then it slows down.

Sing the song again tomorrow, and it stays with you much longer. Every time you practice again, you forget slower than before.

The curve's functional form is still argued. Ebbinghaus fit a logarithmic law; Wixted and Ebbesen's reanalyses favor a power function, and averaging over items or subjects can manufacture a power law out of exponentially forgetting units. Jost's law adds a second regularity: of two traces at equal strength, the older one loses strength more slowly — which is exactly what repeated, spaced resets exploit.

Mechanistically, savings implies that forgetting is often loss of access rather than erasure — a relearnable trace persists below the recall threshold. Interference theory locates the loss in competition from other learning; consolidation accounts locate it in incomplete stabilization. Both predict the observed lever: each successful relearning yields a shallower subsequent slope.

Read as method, Ebbinghaus 1885 is a single-subject design with relearning to criterion, serial-position confounds, and heroic control of conditions — and its shape has survived every replication attempt, most directly Murre and Dros (2015). Savings anticipates modern implicit measures: it registers subthreshold trace strength that recall and recognition both miss.

Modern schedulers formalize the reset. ACT-R's base-level activation decays as a power of time, with each retrieval adding a strengthening event; SM-2, which this Archive runs, assumes a decay whose time constant grows with every successful review. Both encode the Bjorks' distinction: a review raises retrieval strength briefly, but the durable purchase is storage strength — visible as the flattening of the next curve.