Why you forget what AI taught you, and how to keep it

In 1879 a psychologist in Berlin started memorising lists of nonsense syllables (DAX, BOK, YAT) and timing how long it took to learn them again later. He was the only subject. It took him years.

Hermann Ebbinghaus published the results in 1885.¹ He measured "savings": how much less work it took to relearn a list than to learn it the first time. An hour after learning a list, relearning it took about half the original work. A day later, about two thirds. A month later, "fully four fifths", in his words.

The shape is the point. Most of the loss happens in the first day; what survives the first week tends to stay. In 2015 Jaap Murre and Joeri Dros at the University of Amsterdam repeated the whole thing, with one of them spending about 70 hours on nonsense syllables. The curve came back the same, with one twist: it isn't perfectly smooth, and probably jumps upwards at the 24-hour mark.²

Nonsense syllables are a worst case; meaningful material fades more slowly. But the shape holds for nearly everything that's been tested since: vocabulary, facts, procedures, the plot of a film. It's why you can follow a long explanation from an AI perfectly and struggle to repeat any of it a week later.

A clear explanation is a pleasure to read. It also does the thinking for you, and the thinking is what memory is made from. Several studies in the last two years have measured what that costs.

In a randomised trial, 120 undergraduate business students studied the same material, half with ChatGPT and half the usual way. Forty-five days later, 85 of them came back for a surprise test. The ChatGPT group remembered 57.5%; the others 68.5%.³

A larger field experiment with about 1,000 high-school maths students in Turkey found the same pattern in two steps.⁴ With plain ChatGPT open during practice, scores went up 48%. When the AI was taken away for the exam, the same students scored 17% lower than students who never had it. A version built as a tutor, which gave hints instead of answers, largely removed the harm.

Two smaller studies point the same way, with the usual caveats. An MIT Media Lab group recorded brain activity while 54 people wrote essays with an AI, a search engine or nothing; the AI group showed the weakest connectivity and they struggled to quote accurately from the essay they'd just written.⁵ It's a preprint, and small. A survey of 666 people in the UK found heavier AI use went with lower critical-thinking scores, a correlation and not a cause.⁶

None of this means AI is bad for learning. The tutor result says the opposite: AI helps when you still have to do the work. That's the bet Nacre makes. Keep learning from AI, then do the part that makes it stick.

When students are asked how they study, the most common answer by far is rereading. In one survey of 177 college students, 84% listed rereading their notes or textbook, and 55% said it was their main strategy. Testing themselves? 11% listed it. Two students, 1%, named it as their main strategy.⁷

It's easy to see why. The second time through, everything looks familiar, and familiar feels like known. Psychologists call it a fluency illusion. The cleanest demonstration is a 2006 experiment by Henry Roediger and Jeffrey Karpicke.⁸ Students read a short passage either four times, three times with one recall test, or once with three recall tests. Then they were tested five minutes later, or a week later.

After five minutes, reading four times wins. After a week, it loses badly. And when asked to predict how well they'd do, the students who'd read four times were the most confident. The strategy that feels best is the one that fails.

The feeling of knowing is a bad guide to what you'll know next week.

The result in the last chart is called the testing effect, or retrieval practice, and it's one of the most replicated findings in the field. Every time you pull something out of memory, you make it easier to pull out again. Looking at the answer does far less.

Meta-analyses pooling hundreds of experiments put the size at about half a standard deviation, in the lab and in real classrooms.^9,11 Another found practice tests beat "restudying and all other comparison conditions".¹⁰

In 2013 five cognitive psychologists reviewed ten popular study techniques for Psychological Science in the Public Interest, rating each on how well the evidence held up across learners, materials and real settings.¹² Two came out on top. Rereading, highlighting and summarising, the three most people use, came near the bottom.

Reviewing the same thing three times in one evening is cramming. Reviewing it three times over three weeks is spacing, and for long-term memory it wins nearly every time. A 2006 review pulled together 317 experiments from 184 articles; spacing beat massing in almost all of them.¹³

The harder question is how far apart. In 2008 the same group ran it at scale: more than 1,350 people learned obscure facts, reviewed them after gaps of different lengths, and were tested 7, 35, 70 or 350 days later.¹⁴ The best gap depends on how long you want to remember. To remember for a week, it was about 20 to 40% of that week. To remember for a year, about 5 to 10% of the year. Getting the gap right raised recall by 64% overall, an effect size of d = 1.1, which is very large for this kind of research.

One review at the right moment can be worth far more than several at the wrong one. Too early and you've wasted it; the memory didn't need the help. Too late and there's nothing left to strengthen.

Studies like Cepeda's give averages. Real memories don't share a schedule: a fact you found easy fades slower than one you struggled with, and every review changes the curve that follows. A modern scheduler models each item separately.

Nacre uses FSRS, the Free Spaced Repetition Scheduler, an open-source model built by Jarrett Ye and the Open Spaced Repetition community from research first published at KDD 2022.¹⁵ For each question it tracks three numbers: how stable the memory is, how difficult the item is for you, and your chance of recalling it right now. It's tested in public against other schedulers on a dataset of about 727 million real flashcard reviews from 10,000 people; about 350 million of them are used to score each algorithm on how well it predicts whether you'll remember.¹⁶

We ran the numbers from the benchmark's published results against SM-2, the algorithm from 1987 that most flashcard apps were built on. FSRS predicted recall more accurately for 9,967 of the 9,999 people, 99.7%, and its average prediction error was about a third of SM-2's.

Here's a result that surprises most people. When you get something wrong that you were sure about, you're more likely to remember the correction than when you were guessing. Brady Butterfield and Janet Metcalfe called it the hypercorrection effect in 2001,¹⁷ and it's held up in many studies since.¹⁸ The surprise does the work: a confident error grabs attention in a way a shrug doesn't.

A related finding: trying to answer before you've learned something, even when you fail, makes the answer stick better once you see it.¹⁹ Guessing primes the slot the answer goes into. Both effects need the same thing. You have to commit before you see the answer.

Try it.

Testing yourself works without feedback, because recall itself strengthens memory. But for the things you get wrong, feedback is the whole story. In a 2005 study of word pairs, Harold Pashler and colleagues found that showing the right answer after an error raised final recall of those items by 494%.²¹ Feedback after correct answers made little difference.

Robert Bjork's phrase for this is desirable difficulties:²² conditions that slow you down while learning and make what you learn last. Spacing is one. Testing is another. Typing an answer from nothing is harder than picking it from a list, and does more.

"Desirable" is doing a lot of work there. Too hard and you get nothing right, learn little, and quit. In 2019 a team from Arizona, Brown and Princeton worked out the best error rate for a broad class of learning algorithms and got 15.87%, so about 85% right.²³ It comes from maths and simulations of learning machines, not from people, and assumes a simple right-or-wrong task. But it matches what good teachers do by feel: mostly success, with real stretch.

In 1978 Norman Slamecka and Peter Graf showed that words people generate themselves (from a clue like hot: c_ _ _) are remembered better than the same words read.²⁴ In every one of their five experiments, generating beat reading. It's called the generation effect, and a 2007 meta-analysis found it held across decades of studies since.²⁵

Explaining works the same way, one level up. Michelene Chi's group found that eighth-graders prompted to explain a biology text to themselves as they read it gained more understanding than classmates who read it twice.²⁶ Putting an idea in your own words forces you to connect it to what you already know, and the gaps show up.

Practising one kind of problem at a time (blocking) feels efficient. Mixing kinds (interleaving) feels chaotic, and you get more wrong while you practise. In a 2007 maths study by Doug Rohrer and Kelli Taylor, the mixed group scored lower during practice and far higher on the test a week later.²⁷ It was a small study, 18 students, but the effect was large (d = 1.34).

People hate this result. In a 2008 study of learning painters' styles, 78% of participants did better when the paintings were spaced and mixed, and 78% said the bunched-up version worked as well or better.²⁹

Mixing forces you to work out which idea applies, not just how to use it, and that's the skill you need when a question doesn't come labelled. Dunlosky's review rated interleaving moderate utility, with most evidence from maths and category learning.¹²

Katherine Rawson and John Dunlosky put testing and spacing together and gave it a name: successive relearning. Practise until you get each item right, then come back in later sessions and do it again.²⁸ They tested it in real university courses, on the material students were actually being examined on.

One correct answer isn't the finish line. Getting it right again, after a gap, is.

Notice when the studies above measure success. Not right after learning. A week later, a month later, 45 days, a year. An answer you give five minutes after reading something says very little; the same answer a month later says a lot.

That's how Nacre's levels and awards work. They count only what you recall after real gaps, answered from memory inside Nacre. Not answers given in a chat with an AI, not after leaving the app to look something up. So when an award says you know something, it means you still knew it weeks later, without help.

Nacre is the material inside a shell. It's built a layer at a time around a single grain until it's one of the strongest natural materials known. We named the app after it because memory works the same way: each well-timed review adds a layer.

Most ways of studying use one or two of these findings. Rereading and highlighting use none. Flashcard apps use two. Nacre is built on every one.

Everything above is other people's research. It tells us how to build Nacre; it doesn't prove Nacre works. Only our own numbers can do that.

Nacre is invite-only while we test it. During this period we're measuring one thing above all: how much people still recall after real gaps of a week, a month and two months, answered from memory, compared with the 45-day result in section two. We count totals across everyone, never anyone's answers or name. We'll publish the results here, whatever they show.