The Forgetting Curve: What Ebbinghaus Actually Found

What is the forgetting curve?

The important structural fact is the shape, not the numbers. Forgetting is front-loaded. Most of what you lose, you lose early — and the longer a memory survives, the slower it decays from that point. This is why a review tomorrow buys you more than a review in ten minutes, and why the fifth review of a card buys you months when the first bought you days.

What Ebbinghaus actually did

The design is worth understanding because it explains why the numbers are so often misread.

Ebbinghaus wanted memory stripped of meaning. Real words carry associations, and associations vary between people and between words, which would have contaminated the measurement. So he constructed consonant–vowel–consonant syllables with no meaning, learned a list to a criterion of perfect recitation, waited a set interval, and then relearned it to that same criterion.

His measure was savings: the proportion of the original learning effort he no longer needed. If the list took 20 minutes to learn first time and 12 minutes to relearn a day later, he had saved 8 of 20 minutes — 40% savings.

This is not the same thing as remembering 40% of the list. Savings can be substantial when direct recall is at zero: the material leaves a trace that speeds relearning even when you cannot produce any of it on demand. Every article that presents Ebbinghaus's percentages as "how much you remember" has mistranslated the unit.

The numbers everyone quotes

Here are the figures, correctly labelled:

Look at the last three rows. Between one day and one month, savings fall from 34% to 21%. Between twenty minutes and one hour they fall from 58% to 44%. The first hour is more destructive than the following month. That asymmetry is the genuinely useful content of the curve, and it survives every legitimate criticism of the numbers.

Four things the popular version gets wrong

1. Savings are reported as recall

Covered above, and it is the most consequential error. When a productivity article tells you that "you forget 56% of what you learn within an hour", it has taken a relearning-efficiency measure and relabelled it as recall. The real recall figures for meaningful material you understood are much better than the curve suggests.

2. One participant, learning nonsense

Ebbinghaus tested Ebbinghaus. That was reasonable methodology for 1885 and he was scrupulous about it — controlling time of day, rate of recitation, and list length — but a sample size of one tells you nothing about variation between people.

More importantly, nonsense syllables are close to the worst-case material. They have no hooks, no meaning, no connection to anything you already know. Ebbinghaus chose them precisely for that reason. But the practical consequence is that his curve is a lower bound on human memory, not a description of it. Material you understand, that connects to knowledge you already have, decays far more slowly.

3. The chart with the review lines is not his data

The image nearly always attached to this topic — a steep curve, then a series of shallower curves stepping up from review markers — is an illustration of a principle. Ebbinghaus did not produce it. He did study repeated learning and did find that additional repetitions, particularly distributed ones, reduced subsequent forgetting; the modern chart is a stylised composite of that idea and a century of later work.

The principle it depicts is broadly correct. It just isn't a graph of anything anyone measured, and the specific interval markers on it ("review at day 1, day 7, day 30") are invented.

4. It is presented as a law rather than a family of curves

There is no single forgetting curve. There is a different curve for every combination of material, person, encoding quality, and number of prior retrievals. Treating "the" forgetting curve as a fixed constant is what leads people to fixed review ladders, which is exactly the wrong conclusion to draw from it.

What has held up

Murre and Dros, 2015. A careful replication using the original methodology — nonsense syllables, savings measure, a single dedicated participant across many sessions. The resulting curve was broadly consistent with Ebbinghaus's, which is a genuinely impressive result for a 130-year-old self-experiment. They also noted an irregularity around the 24-hour mark, plausibly connected to sleep.

Wixted and Ebbesen, 1991. Examined the mathematical form of forgetting and found a power function fitted better than an exponential across several datasets. The distinction matters: exponential decay means a constant proportional loss per unit time, whereas a power function means the loss rate itself decreases — older memories decay more slowly than young ones, even at equal strength.

Rubin and Wenzel, 1996. Tested over a hundred candidate mathematical functions against more than two hundred retention datasets. No single function won everywhere; logarithmic, power, and exponential-power forms performed consistently well. The honest summary is that forgetting has a characteristic decelerating shape which several functions approximate, not one true equation.

Why some memories stop decaying

This is the most encouraging result in the memory literature and it is almost never mentioned alongside the forgetting curve, presumably because it undercuts the alarm.

Two features predicted who kept their Spanish, and neither was how well they did in the class. What mattered was how many courses they had taken — that is, over how long a period the material had been revisited — and, to a lesser degree, the grade achieved. Time elapsed since the last course mattered enormously for the first few years and then essentially stopped mattering.

The practical reading: the goal of spaced review is not to hold a memory up forever against a downward pull. It is to get material through the first few years of decay often enough that what remains stops eroding. Curves flatten. They just take longer to flatten than any study plan usually accounts for.

Is forgetting decay, or failure to find?

Two accounts compete. Under decay, the trace itself weakens with time until there is nothing left. Under cue-dependent forgetting, the trace persists but the path to it is blocked — usually by interference from other, similar memories competing for the same cue.

The evidence leans hard toward the second. Three observations, none of which decay explains well:

• Savings without recall. Ebbinghaus's own measure makes the point. You can be unable to produce a single item from a list and still relearn it substantially faster than a fresh one. Something survived that recall could not reach.
• Tip-of-the-tongue states. You know the word has three syllables and starts with a plosive. That is not an absent memory; it is a partially accessible one.
• Recovery with a better cue. Material you have "forgotten" often returns instantly given a context, a first letter, or the room you learned it in. Nothing was restored in that instant — the cue simply found it.

Two practical consequences follow, and both are actionable.

First, if you can't recall something, the fault may be the question rather than your memory. A card you keep failing may be perfectly well encoded behind a cue that doesn't reach it — which is why rewriting the front of a failing card so often fixes it where more repetitions did not.

Second, interference is a bigger enemy than time. The material most at risk in your deck is not the oldest; it's whatever most closely resembles something else you're learning. Two similar drug names, two similar particles, two similar case holdings — these decay no faster than anything else, but they become progressively harder to tell apart, which looks identical to forgetting from the inside.

What determines your forgetting rate

Since there isn't one curve, the useful question is what moves yours.

The row worth dwelling on is the last. The forgetting curve is usually framed as a story about time, which makes it feel like an inexorable force. Most of the variance you can actually control is in the rows above it.

How schedulers use the curve

This is the point where the forgetting curve stops being a poster on a classroom wall and becomes machinery. It's worth seeing the actual form, because it is simpler than it sounds:

FSRS models retrievability as R(t) = (1 + f · t/S)^−0.5, where t is days since the last review, S is the card's stability, and f is a constant equal to 19/81. Substitute t = S and the expression collapses to exactly 0.9 — which is the definition of stability doing its job. A card whose stability is 40 days has, by construction, a 90% chance of being recalled 40 days after its last review.

Three things follow directly, and each is a practical decision you can make:

• The exponent is negative and fractional, so this is a power curve, not exponential decay — consistent with Wixted and Ebbesen. Long intervals lose accuracy more slowly than an exponential model predicts, which is why modern schedulers push mature cards further out than older ones did.
• Successful reviews increase S, and increase it more when the review was difficult — the mathematical statement of the desirable-difficulty idea.
• Your retention target is a dial on this curve. Asking for 95% instead of 90% means reviewing every card earlier on its curve, which costs substantially more reviews for a modest accuracy gain.

The details of how S and difficulty are updated, and how the parameters are fitted to your own history, are covered in FSRS vs SM-2.

What to do about it

Five things follow from the curve as it actually is, rather than as it's usually drawn.

1. Review before you've forgotten, not after. The relearning cost of a memory you've genuinely lost is close to the original learning cost. Catching it while it's weak but present is enormously cheaper.
2. But don't review while it's still fresh. A review that costs no effort adds almost nothing. The value is concentrated in retrievals that are hard but successful.
3. The first 24 hours are disproportionate. If you attend a lecture and do nothing until the weekend, you're paying the steepest part of the curve for free. A ten-minute recall attempt that evening is the highest-return study time available to you.
4. Understand first. Ebbinghaus's curve is the curve for meaningless material. Every bit of meaning you attach to something flattens it. This is not a motivational sentiment; it is the largest single difference between his numbers and yours.
5. Stop trying to beat the curve manually. Tracking which of 600 cards is approaching 90% retrievability is not a human task. Hand it to software and spend your attention on the material.

Frequently asked questions

Where this comes from