Spaced Repetition Explained: The Science Behind Long-Term Memory

Published 8 de abril de 2026

In 1885, a German psychologist sat down with lists of nonsense syllables and spent months memorizing them. He tested himself repeatedly, measuring not just whether he remembered, but how much less time it took to relearn a list after different delays.

His name was Hermann Ebbinghaus. The data he published changed how researchers think about memory. It should also change how you study.

TL;DR: Spaced repetition means reviewing material at increasing intervals, timed to just before you'd forget it. Studies show it produces 75–100% better long-term recall than cramming with the same study time. The research goes back 140 years. The algorithms that automate it are now better than ever.

What Ebbinghaus actually found
Why spacing works: the cognitive science
The Leitner system: the original version
From SM-2 to FSRS: how algorithms evolved
What the research says about results
Manual flashcards vs. apps
Where AI fits in
How to start today

What Ebbinghaus actually found

Ebbinghaus spent roughly a year testing his own memory on lists of meaningless letter combinations — "DAX," "BUP," "ZOL," and hundreds more. He measured how much time he saved when relearning a list, which gave him a precise measurement of retention.

The numbers are worth knowing specifically, because "you forget 70% in 24 hours" gets cited everywhere but isn't quite right. Here's what Ebbinghaus actually measured:

20 minutes after learning: 58% retained
1 hour later: 44% retained
9 hours later: 36% retained
1 day later: 33% retained (67% forgotten, not 70%)
6 days later: 25% retained
31 days later: 21% retained

A 2015 replication at the University of Amsterdam confirmed these numbers hold up. One important caveat: Ebbinghaus used meaningless syllables by design, to isolate pure memory from prior knowledge. Retention of material you actually understand is substantially higher.

What the curve shows is the shape of forgetting, not a fixed fate. Most forgetting happens in the first few hours. After that, the curve flattens. And each review resets it. A word reviewed at day one, day six, and day thirty is far more durable than one studied three times in a row on the same afternoon.

Why spacing works

Two mechanisms explain most of the spacing effect.

Retrieval difficulty. When you try to recall something and it takes effort, that effort strengthens the memory more than easy recall does. Reviewing something you studied five minutes ago requires almost no effort — it's still active in working memory. Reviewing something from two weeks ago is genuinely hard. That difficulty is the point. Robert Bjork at UCLA calls these "desirable difficulties" — conditions that feel harder in the short term and produce better long-term retention.

The corollary: passive review (rereading notes, re-watching lectures) feels productive but doesn't trigger this mechanism. You're recognizing, not retrieving.

Memory reconsolidation. Every time you access a memory, you temporarily destabilize it. The brain then restores it in a more durable form. This is why long-term memories aren't static — they're updated and strengthened each time you use them. A memory retrieved multiple times over months is structurally more stable than one encoded once, however deeply.

These two mechanisms work together. Spacing increases retrieval difficulty, which triggers deeper reconsolidation. The timing matters because reviewing too soon bypasses the first mechanism entirely.

There's also a third effect worth knowing about: the illusion of fluency. When you reread your notes and they feel familiar, your brain interprets that familiarity as "I know this." It doesn't. Familiarity and retrievability are different things. You can recognize the right answer when you see it and still fail to produce it under exam conditions. Spaced retrieval practice tests actual retrievability — whether you can generate the answer unprompted — rather than just recognition.

This is why cramming works the morning of an exam and fails three weeks later. The information is temporarily accessible. The spacing effect creates durable access.

The Leitner system: the original version

Sebastian Leitner was a German science journalist who published a practical spaced repetition method in his 1973 book, So lernt man lernen ("How to Learn to Learn"). No software, no algorithm — just boxes and paper cards.

The system is straightforward. You have five boxes. Every card starts in Box 1, reviewed daily. Correct answer: the card moves to Box 2, reviewed every two days. Another correct answer sends it to Box 3 (every four days), Box 4 (every eight days), Box 5 (every sixteen days). Wrong answer at any box: back to Box 1.

The elegance is that the system self-organizes. Cards you know well naturally require less time. Cards you keep getting wrong keep coming back. Your study time concentrates on what needs it, without you having to manage that consciously.

The limits are equally clear. Fixed intervals don't adapt to how your memory actually works — a card you barely squeezed out is treated identically to one you recalled instantly. The system also becomes unwieldy once you're managing thousands of cards across multiple subjects.

That's what software solved.

From SM-2 to FSRS: how algorithms evolved

Piotr Wozniak was a Polish university student in the mid-1980s who wanted to retain English vocabulary and biology facts long-term. He ran his own experiments, tracking review intervals manually. On December 13, 1987, he implemented the result as SuperMemo, the first spaced repetition software.

The version that spread was SM-2. It works like this:

First review: 1 day after initial study
Second review: 6 days after the first
Subsequent reviews: previous interval multiplied by an "easiness factor" (EF)
EF starts at 2.5 for all cards and adjusts down for harder cards (minimum 1.3)
After each review, you rate your recall on a 0–5 scale; EF recalculates accordingly

Wozniak's own usage data showed roughly 89–90% retention — which is the target the algorithm is designed to hit. SM-2 is still what powers Anki today, and it works well enough that "well enough" undersells it. It has helped hundreds of thousands of medical students pass licensing exams.

The limitation: it applies the same formula to everyone. Your forgetting curve isn't the same as mine. SM-2 assumes it is.

FSRS (Free Spaced Repetition Scheduler) fixes this. Developed by Jarrett Ye and first published in 2022, it was integrated into Anki in November 2023 and is now the recommended algorithm there. The underlying model tracks three variables per card: Difficulty (how inherently hard the card is), Stability (how long the memory is expected to last), and Retrievability (current probability of recall). It fits 21 parameters to your personal review history using machine learning.

The practical result: FSRS produces roughly 20–30% fewer reviews for the same retention target compared to SM-2. It also handles missed reviews better — if you fall behind for a week, SM-2 treats many cards as essentially forgotten and schedules a flood of reviews. FSRS models what actually happened to your memory during that time and adjusts accordingly.

To put that in terms of what a review schedule actually looks like: if you learn a new Spanish word today and rate it as easy, FSRS might schedule the next review in 4 days, then 12 days, then 35 days, then 90 days. A difficult word you keep forgetting might get reviews every 1–3 days for weeks before the intervals start expanding. The algorithm adapts to each card individually, based on your actual response history, rather than applying the same progression to everything.

For most people, the algorithm difference isn't what determines results. Actually doing the reviews is. But if you're committed to a years-long study schedule — medical school, a new language, bar exam prep — the efficiency difference adds up.

What the research says about results

Spaced repetition has been studied a lot. The results keep pointing the same direction.

Dunlosky et al. (2013) evaluated 10 common study strategies — highlighting, summarising, rereading, practice testing, distributed practice, and several others — against demanding criteria: long-term retention, transfer of learning, applicability across subjects and age groups. Distributed practice and practice testing were the only two rated "high utility." Rereading and highlighting, the techniques most students lean on, were rated "low utility."

Cepeda et al. (2006) analyzed 839 assessments from 317 experiments and confirmed the spacing effect across a vast range of material and populations. The key finding: optimal spacing gaps aren't fixed — they scale with how long you need to retain the material. A 2008 follow-up with 1,350+ participants got specific: for material you need in one week, the optimal review gap is roughly 20–40% of that delay (one to three days). For material you need to retain for a year, the optimal gap is 5–10% of the delay (three to five weeks). Most studying happens on schedules far denser than this.

Kornell & Bjork (2008) ran a direct comparison of spaced versus massed learning. Spaced learners scored 61%, massed learners scored 35% on final tests — roughly 75% better performance from the same total study time, just distributed differently.

Bahrick et al. (1993) studied Spanish vocabulary retention over nine years with different spacing schedules. The finding that stands out: 56-day spacing over 13 sessions produced the same long-term retention as 14-day spacing over 26 sessions. Wider gaps, half the sessions.

The effect holds across languages, medicine, history, and mathematics — essentially everywhere researchers have looked for it.

One thing the research makes clear that often gets lost: the spacing effect scales with the retention interval. If you're studying for an exam next week, cramming-adjacent schedules are actually close to optimal — you need the material in a few days, so large spacing gaps would be counterproductive. Where spaced repetition creates the largest advantage is when you need to retain material for months or years. The benefits compound over time in a way that front-loaded studying simply can't match.

Manual flashcards vs. apps

Physical flashcards with a Leitner box work. Writing cards by hand encodes material more deeply for many people, and there's something to be said for a study system that doesn't involve a screen. If you're the type who learns better when you write things down, a physical system can produce better encoding at the creation stage, even if it's less optimal at the scheduling stage.

The practical limits show up fairly quickly. Manual tracking across multiple subjects gets complicated fast. Fixed intervals don't adapt to how your memory actually works. You have no data on performance over time — no way to see that you've been consistently forgetting a specific category of material. For a few hundred cards on one subject, physical is fine. For thousands of cards across a year of studying, it becomes its own maintenance job.

Apps handle scheduling automatically. You review a card, rate your recall, the algorithm adjusts. The cards you're likely to forget show up more often. The ones you know well fall into the background until they need refreshing.

The main risk with apps is the same as with any tool: spending more time configuring than studying. Anki's add-ons, card templates, and tag systems can absorb hours that should go into actual reviews. A simple Leitner box you actually use beats a meticulously organized Anki deck you maintain instead of review.

Where AI fits in

The bottleneck in spaced repetition has never been the reviewing. It's making the cards.

Creating a quality Anki deck from a textbook chapter takes real time — probably an hour for a dense chapter, more if you're being thorough. Most people who start flashcard systems and abandon them do so at this stage, not because reviewing is hard, but because setup feels endless before you've even started.

AI removes that bottleneck. Upload a lecture PDF, paste in notes, or share a YouTube lecture link, and a tool like Quizgecko's flashcard generator produces a complete flashcard set in under a minute, with spaced repetition scheduling built in. The 90-minute setup cost drops to about ten.

The cards aren't always perfect — you'll want to edit some, and that editing is itself useful active engagement with the material. But the barrier to starting a new deck disappears.

Beyond flashcards: an AI quiz generator can produce multiple-choice, short answer, and fill-in-the-blank questions from the same material. Dunlosky's review found distributed practice and retrieval testing to be the two most effective techniques in the literature — combining them produces both benefits at once.

If you want to generate a full study package from any document or video, Quizgecko's AI study guide maker builds structured summaries alongside the flashcards, covering the "understand it first" part of learning before the spaced repetition begins.

How to start

The practical steps are simpler than most guides make them.

First, decide what actually needs to stick long-term. Not everything does. For material you need tomorrow and never again, cramming is fine. Spaced repetition is for vocabulary, anatomy, pharmacology, legal concepts, historical formulas — things you need accessible for months.

Then break that material into discrete, answerable facts. One card, one question, one answer. "Explain the krebs cycle" is not a good flashcard. "What is the net ATP yield of the krebs cycle per turn?" is. The smaller and more specific, the cleaner the review.

Review daily, but keep sessions short. Ten to fifteen minutes daily beats two hours on Sunday. The system needs regularity more than volume. Missing a day occasionally is fine; missing a week lets cards pile up into a backlog that feels punishing to clear.

When you review, cover the answer and actually try to retrieve it before checking. This is the mechanism that makes spaced repetition work. Skipping the retrieval step — reading through cards without generating answers — turns it into rereading. Rereading is low utility.

A concrete example: a nursing student studying pharmacology might create cards for 200 drug names, mechanisms, and dosages. In week one, all 200 are new. She reviews every day. By week three, maybe 60 of those cards have been answered correctly enough times to move to longer intervals — she sees them every five days, then every two weeks. The other 140 she's still struggling with appear daily or every other day. By month two, the easy cards barely require any time. Her daily review session takes 10–15 minutes instead of 45, and it's concentrated entirely on her weak points.

That's what the algorithm does automatically. Without it, you either review everything constantly (inefficient) or stop reviewing things you think you know (risky — memory fades faster than you expect).

The payoff comes slowly and then obviously. Material reviewed in September still accessible in February. Vocabulary that required three readings to stick, now permanent after five spaced reviews. The benefit is most visible on long timescales — which is exactly why cramming feels effective in the short term. It works for tomorrow's exam. Spaced repetition is for the exam in six months, and the one after that.

Ebbinghaus mapped the shape of forgetting in 1885. The shape hasn't changed. What's changed is how precisely we can time interventions, and how much of the work the tools can do for you. If you want to generate flashcards from your own material and let the scheduling handle itself, that's a reasonable place to start. The AI quiz generator adds retrieval testing on top of it.

The rest is showing up for the reviews.