The forgetting curve, explained: why you lose most of what you learn, and how to keep it

Learn something today and, if you never touch it again, most of it will be gone within days. That is not a personal flaw. It is one of the oldest quantified findings in psychology, and it has a shape: the forgetting curve.

Understanding that shape changes how you study, because the curve is not just a description of failure. It is a map of exactly when a review does the most good.

The short answer

The forgetting curve describes how memory retention drops over time without review: steeply at first, then more slowly. Hermann Ebbinghaus first measured it on himself in the 1880s by memorizing nonsense syllables and testing how much relearning they needed later. His numbers suggest you lose roughly half of new, unconnected material within about an hour, and around 70 percent within a day. After that early cliff, the decay flattens.

Two things stop the slide: retrieving the memory, and retrieving it again at growing intervals. Each successful recall resets the curve and makes the next decay slower. That is the entire logic of spaced repetition.

Where the curve comes from

Ebbinghaus was the first person to study memory experimentally rather than philosophically. He invented lists of meaningless syllables ("WID", "ZOF") so that prior knowledge could not help, memorized them until he could recite them perfectly, and later measured how much effort it took to relearn them. The savings in relearning time became his measure of what remained.

The result was a curve that falls fast and then levels off, well approximated by an exponential decay. The remarkable part is not the 1885 experiment; it is that the finding held. In 2015, Murre and Dros repeated the original procedure with modern controls and reproduced the shape of Ebbinghaus's curve in detail.

What the pop version gets wrong

The forgetting curve gets quoted a lot, and a few distortions have crept in.

• The exact percentages are not universal. "You forget 70 percent in 24 hours" comes from one man memorizing nonsense syllables. Meaningful material, connected to things you already know, decays more slowly. The shape generalizes; the specific numbers do not.
• Forgetting is not malfunction. The steep early drop is your brain triaging. Most of what passes through your attention genuinely does not need to be kept, so the default is to let it fade and keep what shows up again. The curve is an efficiency feature you can exploit, not a bug you can wish away.
• The curve is not fixed. Every review changes it. This is the part most summaries skip, and it is the part that matters.

The part that matters: resetting the curve

Retrieve a memory just before it slips away and two things happen. Retention jumps back up, and the new forgetting curve decays more slowly than the last one. Do this a few times and the intervals between necessary reviews stretch from a day, to several days, to weeks, to months.

The order of operations matters, though. The reset is much stronger when the review is a retrieval, an attempt to produce the answer from memory, rather than a re-exposure. Re-reading the material also bumps retention, but weakly and briefly, and it feels far more effective than it is. That gap between felt learning and real learning is why your notes don't stick, and the fix, active recall, is a topic of its own.

So the practical recipe is: retrieve, at growing intervals, timed to land just before you would forget. The spacing effect behind this is one of the most replicated findings in learning research, documented across hundreds of studies.

From curve to schedule

For a handful of facts you could manage those intervals by hand. For hundreds of flashcards across several subjects, each on its own decay clock, you cannot. This is exactly the problem spaced repetition software solves: track every item's curve and schedule its review for the moment retention is about to dip below a target.

How well the software does that depends on how well it models your forgetting. Early algorithms like SM-2 used a fixed 1980s formula. Modern ones, like FSRS, fit a memory model to hundreds of millions of real reviews and predict each card's retention directly. The difference is measurable, and we covered it in detail in FSRS vs SM-2: roughly the same retention for 20 to 30 percent fewer reviews.

In other words: the forgetting curve is the problem statement, retrieval is the mechanism, and a modern scheduler is the automation. None of the three is optional if you want what you learn in January to still be there in June.

Frequently asked questions

How fast do you actually forget new information? For arbitrary, unconnected material, Ebbinghaus's data suggests about half is gone within an hour and around 70 percent within a day. Meaningful, well-connected material fades slower, but the shape, fast early loss then a long tail, is the same.

Is the forgetting curve scientifically valid? Yes. Beyond a century of compatible results, the original experiment was replicated in 2015 by Murre and Dros, who reproduced the curve with modern methodology.

Can you stop forgetting entirely? No, and you would not want to. What you can do is make the decay slow enough not to matter, by spacing retrievals so each one lands just before the memory slips.

What is the ideal time to review something new? Soon after learning (within a day), then at stretching intervals: a few days, a week, a few weeks. The precise optimum depends on the material and on you, which is why algorithms that fit your actual review history outperform fixed rules.

Does the forgetting curve apply to skills? Skills decay too, though usually more slowly than facts, and they respond to the same treatment: spaced, effortful practice rather than re-watching the tutorial.

The bottom line

The forgetting curve says that losing most of what you learn is the default, not the exception, and that the loss is front-loaded into the first hours and days. It also says the default is negotiable: each well-timed retrieval resets the curve and flattens the next one.

That loop is tedious to run by hand, which is why we built it into the notebook itself. Write a page in Anti-Agent on whatever you are learning, and the page schedules its own returns with FSRS: flashcards, dialogues, and exercises that arrive just as the curve says you are about to lose the thread.