Quizlet Spaced Repetition

Introduction

Here is a number that should bother anyone who studies with flashcards: 95 percent of Quizlet users study a given set for four days or fewer [1]. That statistic comes not from a critic but from Quizlet's own engineering team, published in a 2017 technical blog post written by machine-learning engineer Shane Mooney. It raises an uncomfortable question. Quizlet spaced repetition is a phrase that millions of students type into Google every month, expecting that the world's most popular flashcard platform uses the same brain science that researchers have validated for over a century. But what if the algorithm behind Quizlet does something fundamentally different from the spaced repetition systems described in textbooks? What if the phrase itself is misleading?

To answer that, the story has to go back. Not to Silicon Valley. Not even to the invention of computers. It starts in a cramped apartment in Berlin in 1879, where a young German philosopher sat down with a metronome, a stack of nonsense syllables, and a plan to measure something no one had ever measured before: the speed at which the human mind forgets [2].

The Man Who Memorized Nonsense

Hermann Ebbinghaus was not a neuroscientist. In 1879 there was no such field. He was a philosopher at the University of Berlin with a radical idea: memory could be measured with the same rigor that physicists applied to light and gravity. To test this, he needed material that carried no prior associations. Real words would not work. A German speaker already "knows" the word Tisch. Any retention of that word would be contaminated by decades of prior exposure. So Ebbinghaus invented a new unit of memory research: the nonsense syllable. Consonant-vowel-consonant combinations like WID, ZOF, BOK, and DAX. He made 2,300 of them [3].

Then he became his own experimental subject.

For five years, Ebbinghaus sat alone, pacing himself with a metronome, reading lists of syllables aloud until he could recite each list perfectly twice. He recorded how long each list took to learn, waited precise intervals (twenty minutes, one hour, nine hours, one day, two days, thirty-one days), and then relearned the same list. The key measurement was not whether he remembered the syllables. It was how much time he saved when relearning them compared to the original session. He called this the "savings method," and it was brilliant because it captured traces of memory too faint for conscious recall. A list that takes ten minutes to learn originally but only six minutes to relearn has a savings score of 40 percent.

The data he collected, published in 1885 as Über das Gedächtnis, produced the forgetting curve: a sharp initial drop in retention (roughly 58 percent lost within twenty minutes, 66 percent within twenty-four hours), followed by a gradual leveling off [4]. The curve had an implication that Ebbinghaus himself recognized. If forgetting follows a predictable mathematical function, then review could be timed to intercept forgetting at precisely the moment a memory is about to disappear. Review too early, and the effort is wasted on material already solid. Review too late, and the memory is gone. The trick is hitting the sweet spot.

That insight is the seed from which all spaced repetition grew. But it would take exactly one hundred and two years before anyone built a computer program to automate it.

What does this mean for students today? Every flashcard app, every study schedule, every "review before you forget" notification traces back to those afternoons in Berlin. The forgetting curve is not a metaphor. It is a measurable, replicable biological process, and in 2015, Dutch psychologist Jaap Murre replicated Ebbinghaus's original experiment with remarkable fidelity, confirming both the curve's shape and an unexpected detail: a small upward jump in retention at the twenty-four-hour mark, likely caused by overnight sleep consolidation [4].

From Flash Cards to Five Boxes

The century between Ebbinghaus and the first computer algorithm was not empty. In 1939, Herbert Spitzer tested 3,605 sixth-graders in Iowa schools, giving them reading passages and then recall tests at various spaced intervals [5]. Spitzer's study was the first large-scale educational demonstration that spaced review beats cramming. But it was largely forgotten for decades. Ebbinghaus had the idea; Spitzer proved it worked in classrooms. Neither had a practical system for individual learners.

That system came from an unlikely source. In 1967, linguist Paul Pimsleur proposed "graduated-interval recall" for language learning: review a new word after five seconds, then twenty-five seconds, then two minutes, then ten minutes, scaling up to months and years [6]. His intervals followed a rough geometric ratio of five. The Pimsleur method is still sold commercially today, and it works, but it was designed for audio courses, not self-study.

The physical system that millions of students would eventually recognize arrived in 1972. Sebastian Leitner, a German science journalist, published So lernt man lernen ("How to Learn to Learn") and introduced the Leitner box. The idea was elegant. Five cardboard compartments. All new flashcards start in Box 1. Get a card right, and it moves to Box 2. Get it right again from Box 2, and it moves to Box 3. Get it wrong at any point, and it drops back to Box 1. Each box is reviewed at a longer interval: Box 1 every day, Box 2 every three days, Box 3 every week, and so on. Correct answers earn longer rest periods. Wrong answers restart the cycle [7].

The Leitner system was a categorical approximation of the spacing effect. It did not calculate precise intervals. It used boxes, not equations. But it captured the core principle: successful recall should lead to longer gaps before the next review, and failure should reset the gap to zero. This is a threshold that matters. Everything called "spaced repetition" after Leitner either refines his insight with mathematics or abandons it entirely. And that distinction is exactly where the Quizlet story gets interesting.

The Algorithm That Changed Everything

In 1985, a Polish university student named Piotr Woźniak did something nobody else had bothered to do: he tracked his own forgetting with mathematical precision. Using paper notebooks, he recorded exactly when he reviewed each fact and whether he got it right. From this data, he built tables of optimal review intervals. He called the system SuperMemo. The paper-based version (SM-0) used fixed intervals: one day, seven days, sixteen days, thirty-five days. It worked, but it was tedious [8].

On December 13, 1987, Woźniak wrote the code that would quietly reshape how millions of people study. SM-2, implemented in Turbo Pascal on an IBM PC, was the first computer algorithm to schedule flashcard reviews automatically based on individual performance [9]. The math was simple. Each card carried an "easiness factor" (EF), initialized at 2.5. After each review, the learner rated their recall on a scale from 0 to 5. The algorithm updated EF using this formula:

EF' = EF + (0.1 - (5 - q) × (0.08 + (5 - q) × 0.02))

Here, q is the quality rating. The factor could never drop below 1.3. Intervals were calculated by multiplying the previous interval by EF. First review: one day. Second review: six days. Every review after that: the previous interval times the easiness factor. A card with EF = 2.5 would be scheduled at intervals of roughly 1, 6, 15, 38, 95, and 237 days.

SM-2 had flaws. The easiness factor conflated item difficulty with memory stability. A lapse (forgetting a card) reset the interval to one day, punishing the learner harshly for a single mistake. Over time, this created what Anki users would later call "ease hell," where difficult cards accumulated punishingly short intervals that never recovered. But SM-2 was the first algorithm that actually worked for individual, per-card scheduling across days, weeks, and months. It is the ancestor of almost every spaced repetition system in use today.

What does this mean practically? When a student opens Anki or any FSRS-based app and sees a card scheduled for review in 23 days, that number traces directly to Woźniak's 1987 code. The formula has been improved, but the architecture (per-card state tracking with computed intervals) has not been replaced.

What Quizlet Actually Does

Now the critical question. Quizlet was founded in October 2005 by Andrew Sutherland, then a high school student. It grew into the world's most popular flashcard platform, with nearly a million daily active users by 2017. Somewhere along the way, "Quizlet spaced repetition" became a search query that millions of students typed, assuming the platform used the same kind of per-card interval scheduling as SuperMemo or Anki.

It does not.

In 2017, Shane Mooney published a detailed technical account of Quizlet's Learn mode on the company's engineering blog [1]. The architecture he described is fundamentally different from SM-2 or FSRS. Quizlet's system is a logistic regression model trained on approximately 1.5 million sampled answers from its database. The model predicts the probability that a learner will correctly recall a given term right now, based on features like previous correctness, time since last answer, and question type. Terms with the lowest predicted recall probability are shown first.

This is adaptive within-session ordering. Not cross-session interval scheduling.

The distinction matters. SM-2 and FSRS maintain a per-card state (stability, difficulty, easiness factor) that persists between sessions and computes a specific future date for review. Quizlet's logistic regression ranks terms within a single study session by predicted recall probability, then resets when the session ends. There is no per-card stability variable. There is no computed interval saying "review this card again in 14 days."

Mooney himself acknowledged the design constraint honestly. He wrote that 95 percent of Quizlet users study a set over four days at most, which was "still too few for the Long-Term Learning algorithm to be effective." Quizlet had briefly offered a traditional spaced repetition feature called Long-Term Learning, but most users never engaged with it because they were cramming for a test tomorrow, not studying for a medical board exam six months away. So Quizlet built an algorithm optimized for the way students actually used the platform: short, intensive sessions with no long-term scheduling.

A later update by Quizlet engineer Karen Sun introduced a Recurrent Power Law memory model [10]. This model was closer to the theoretical framework behind FSRS, incorporating time-dependent forgetting into predictions. But the publicly documented implementation still operated within sessions, not across them.

Methodological note: the 1.5 million answer sample Mooney described was drawn from Quizlet's database of hundreds of millions of study sessions. The logistic regression was chosen for portability. Mooney explained that it produced a simple model that could be transcribed into a JavaScript module for use on the website and compiled into both iOS and Android apps. Engineering simplicity was an explicit design goal.

What does this mean for a student choosing a study tool? If the goal is cramming for a test this week, Quizlet's session-scoped approach is reasonable. The algorithm ensures weak terms get more attention during the current session. But if the goal is retaining medical terminology for board exams six months from now, or building long-term vocabulary in a foreign language, Quizlet's architecture is not designed for that task. That is not a deficiency. It is a design decision that matches how most Quizlet users actually study.

The Algorithm That Replaced SM-2

While Quizlet was building for crammers, a Chinese computer science undergraduate was building for the opposite use case. In 2022, Junyao "Jarrett" Ye, then a student at the Harbin Institute of Technology in Shenzhen, published a paper at ACM KDD (one of the top conferences in data mining) proposing a fundamentally new approach to spaced repetition scheduling [11]. The algorithm was called FSRS, the Free Spaced Repetition Scheduler.

FSRS replaced SM-2's single easiness factor with three variables: Difficulty (D), how inherently hard the card is on a scale of 1 to 10; Stability (S), the interval in days at which the probability of recall equals 90 percent; and Retrievability (R), the current estimated probability that the learner can recall the card right now.

The forgetting curve in FSRS is modeled as a power law:

R(t, S) = (1 + F × t/S)^(-c)

Here, t is the time elapsed since the last review, S is stability, and F and c are fitted parameters. In FSRS-6, released in 2025, even c is trainable per user, allowing the shape of the forgetting curve to vary across individuals. Some people forget steeply. Others forget gradually. FSRS-6 adapts to both [12].

The default parameters for FSRS-6 were trained on approximately 700 million reviews from about 10,000 Anki users who volunteered their anonymized review logs (the anki-revlogs-10k dataset). This is orders of magnitude more data than any previous SRS algorithm used for calibration.

In November 2023, Anki version 23.10 adopted FSRS as its default scheduler, replacing SM-2 after more than fifteen years [13]. The Open SRS Benchmark, an independent comparison maintained on GitHub, evaluated FSRS-6 against SM-2 across approximately 9,999 Anki collections containing roughly 350 million review predictions. The result: FSRS-6 had lower prediction error (log-loss) than SM-2 in 99.6 percent of tested collections [14]. Simulation data from the benchmark project suggests that students using FSRS need 20 to 30 percent fewer reviews to maintain the same retention rate [15].

Methodological note: the 20 to 30 percent efficiency figure is based on simulation, not a controlled trial with real students. The direction of the effect is strong, but the precise magnitude should be treated as an estimate. SM-2 was also not designed to predict probabilities. The benchmark forces a probability interpretation onto its output, which may not be entirely fair to Woźniak's original design intent.

Why Spacing Works: The Molecular Evidence

The behavioral evidence for spaced repetition is overwhelming. A 2006 meta-analysis by Cepeda and colleagues synthesized 839 effect-size assessments from 317 experiments across 184 articles and concluded that distributing practice across time consistently beats massing practice into a single session [16]. In 2013, Dunlosky and colleagues reviewed ten popular study techniques for the journal Psychological Science in the Public Interest and rated only two as "high utility" across all ages, materials, and outcome measures: distributed practice and practice testing [17]. Everything else, including highlighting, rereading, and summarizing, received lower ratings.

But why does spacing work at the level of neurons?

The answer begins in 1973, in a laboratory at the University of Oslo. Tim Bliss and Terje Lømo were recording electrical activity from the brains of anesthetized rabbits. They stimulated a pathway called the perforant path, which connects to the dentate gyrus region of the hippocampus, a brain structure essential for forming new memories. When they delivered a burst of high-frequency electrical stimulation (a tetanus), the response of neurons on the receiving end became stronger. And the strengthening persisted for hours [18].

They had discovered long-term potentiation, or LTP: the strengthening of synaptic connections through repeated stimulation. LTP is the cellular mechanism most widely accepted as the basis of learning and memory.

But LTP has two phases, and the distinction between them is what makes spacing necessary at the molecular level. Early-phase LTP (E-LTP) lasts one to three hours and requires no new protein synthesis. It depends on existing receptors being modified and more receptors being inserted into the synapse. Late-phase LTP (L-LTP) lasts hours to days and requires the cell to transcribe new messenger RNA and synthesize new proteins [19]. The protein CREB (cAMP response element-binding protein) acts as a molecular switch. When CREB is activated, it triggers gene expression that produces the proteins needed to structurally remodel the synapse and make the strengthening permanent [20].

Here is the critical connection to spacing. The protein-synthesis window for L-LTP takes roughly one to two hours to fully engage after stimulation. Massed repetitions (cramming) saturate E-LTP but never trigger L-LTP because the cell does not have time to complete the transcription-translation cycle between stimuli. Spaced repetitions, with gaps of hours or longer, allow each retrieval event to trigger a fresh round of protein synthesis, building progressively stronger synaptic architecture.

BDNF (brain-derived neurotrophic factor) plays a central role in this consolidation process. Bramham and Messaoudi described BDNF as "a trigger for protein synthesis-dependent late-phase LTP, a process referred to as synaptic consolidation" [21]. Without BDNF signaling, the molecular machinery of long-term memory simply does not engage.

What does this mean in plain terms? Cramming builds memories out of sand. Spaced review builds them out of concrete. The difference is not effort or willpower. It is protein synthesis.

The Overnight Factory

There is another reason spacing works better when reviews span at least one night of sleep. During slow-wave sleep (the deepest stage of non-REM sleep), three brain rhythms synchronize in a precise temporal sequence. Slow oscillations at about 0.75 Hz sweep across the neocortex. During the depolarizing "up" phase of each slow oscillation, thalamocortical sleep spindles (12 to 15 Hz) fire. And nested within those spindles, the hippocampus generates sharp-wave ripples at 100 to 250 Hz [22].

This triple-nested rhythm is not random electrical noise. It is a replay mechanism. The hippocampus reactivates memory traces acquired during the day and transmits them, via the spindle-ripple coupling, to the neocortex for long-term storage. This process is called systems consolidation, and it was first theorized by McClelland, McNaughton, and O'Reilly in 1995 [23]. Their complementary learning systems theory proposed that "memories are first stored via synaptic changes in the hippocampal system" and that "neocortical synapses change a little on each reinstatement."

Born and Wilhelm formalized this as the active systems consolidation hypothesis in 2012, arguing that sleep is not a passive period of reduced forgetting but an active process of memory reorganization [24]. Rasch and Born's 2013 review in Physiological Reviews provided the most complete synthesis of the evidence: slow-wave sleep consolidates declarative (fact-based) memories, while REM sleep may preferentially consolidate procedural and emotional memories [25].

Methodological note: much of the evidence for sleep-dependent memory consolidation comes from studies using polysomnography (recording brain waves, eye movements, and muscle activity during sleep) combined with memory tasks before and after sleep. Some studies use targeted memory reactivation, where sounds or smells associated with learned material are presented during sleep to boost consolidation of specific memories.

This is directly relevant to the spacing versus cramming debate. A spaced review schedule that spans at least one sleep cycle gives the brain's overnight consolidation machinery time to process each retrieval event. Murre and Dros's 2015 replication of the forgetting curve found a measurable upward jump in retention at the twenty-four-hour mark, exactly where overnight consolidation would be expected to intervene [4]. Cramming, by definition, skips this process entirely.

The Testing Effect: Why Retrieval Beats Rereading

Spacing is only half the equation. The other half is what happens during each review. In 2008, Jeffrey Karpicke and Henry Roediger published a study in Science that produced a result many educators found counterintuitive [26]. They had students learn Swahili-English word pairs. Some students studied the pairs repeatedly. Others were tested on them repeatedly. The critical finding: once a pair had been correctly recalled, repeated studying had no measurable effect on delayed recall. But repeated testing produced a large positive effect.

Even more striking: students' own predictions of how well they would perform were uncorrelated with actual performance. Students who studied without testing felt confident. Students who tested without extra study felt uncertain. But on the actual test one week later, the testing group dramatically outperformed the studying group. Karpicke and Roediger's earlier work had established the same pattern: testing is a "memory modifier," not merely a "memory assessor" [27].

This is the testing effect (also called retrieval practice), and it explains why flashcard-based spaced repetition works better than simply rereading notes on a schedule. The act of trying to recall an answer, and succeeding or failing, changes the memory trace in ways that passive review does not. Every flashcard app that asks a question and waits for an answer before showing the correct response is using this principle. Apps that only show information without requiring retrieval are missing the mechanism that makes spaced repetition effective.

A 2021 systematic review confirmed the effect holds across K-12 and university classrooms in real-world settings [28].

What the Latest Research Shows

The evidence base for spaced repetition continued to grow through 2024 and 2025. In medical education, where the volume of material makes efficient study methods a practical necessity, several studies tested spaced repetition against traditional study methods under controlled conditions.

Gilbert and colleagues (2023) conducted a cohort study at a medical school and found that students who regularly used Anki as a spaced repetition tool scored 6.2 to 10.7 percent higher on standardized exams than students who did not [29]. Durrani and colleagues (2024) ran a study in Pakistan specifically testing spaced repetition for clinical problem-solving in paediatrics, with positive results [30]. And in 2025, Vagha and colleagues at Jawaharlal Nehru Medical College in Wardha, India, conducted a quasi-experimental study with 90 final-year medical students. The intervention group used custom digital flashcards reviewed at intervals of 1, 3, 7, 14, and 28 days. Post-test scores were 16.24 for the intervention group versus 11.89 for the control group (p < 0.0001), and over 90 percent of intervention students reported improved retention [31].

Methodological note: the Vagha study was a single-center quasi-experiment with 45 students per group, no blinding, and no long-term follow-up beyond the study period. The effect is statistically significant but the sample is small and the design is not randomized.

The strongest evidence comes from meta-analysis. Maye and colleagues published a systematic review and meta-analysis in The Clinical Teacher in 2026, pooling data from 14 studies totaling 21,415 learners [32]. Their findings: a standardized mean difference of 0.78 (95% CI: 0.56 to 0.99, p < 0.0001) in favor of spaced repetition compared to standard study techniques. In educational research, a standardized mean difference (the size of the effect measured in standard deviations) of 0.78 is considered a large effect. For context, the average educational intervention produces an effect size around 0.40 [17].

These findings are consistent across the broader literature. Baddeley and Longman demonstrated the spacing effect for motor learning back in 1978, training postal workers to type on alphanumeric keyboards. The group trained for one hour per day learned most efficiently. The group trained for two two-hour sessions per day performed worst, despite accumulating the same total practice time [33].

The Counterargument: When Spacing Fails

Not all evidence is uniformly positive, and honest science requires acknowledging the limitations.

First, the 20 to 30 percent efficiency gain attributed to FSRS over SM-2 is based on simulated review schedules applied to historical data, not on a randomized controlled trial where real students used one algorithm versus the other and took the same exam [12]. The direction of the effect is well-supported, but the precise magnitude may not generalize to all populations.

Second, spaced repetition is most effective for factual recall (vocabulary, definitions, dates, formulas) and less clearly effective for conceptual understanding, problem-solving, or creative thinking. Dunlosky's 2013 review noted that distributed practice and testing have their strongest effects on "retention of facts and concepts" rather than on "transfer" (applying knowledge to new situations) [17]. A student preparing for a multiple-choice anatomy exam will benefit enormously. A student preparing for an essay on political philosophy may benefit less.

Third, the exponential-versus-power-law debate about the shape of the forgetting curve remains unresolved. FSRS uses a power-law model. SM-2 implicitly assumes multiplicative exponential decay. The Rubin and Wenzel (1996) analysis found that over 200 different mathematical functions could fit forgetting data reasonably well. The "correct" shape may depend on the type of material, the learner, and the timescale being measured.

Fourth, Quizlet's design choice is not irrational. For a user who will study a set for three days and take a test on the fourth day, a cross-session scheduling algorithm is irrelevant. The sessions are too close together for interval-based scheduling to differentiate itself from Quizlet's within-session optimization. Quizlet built for its actual users, not for an idealized version of how students "should" study. Whether that is a limitation or a pragmatic engineering decision depends on the individual learner's goals.

What Comes Next

The field is moving fast. SuperMemo released SM-20 in 2026, which Woźniak describes as the first version where all parameters are computed by machine learning rather than hand-tuned heuristics [34]. FSRS-7 is in development, targeting fractional interval lengths (scheduling in hours rather than whole days). The SuperMemo API launched in early 2026, running the SM-20 algorithm as a cloud service. And AI-enhanced scheduling, where language models generate flashcards while spaced repetition algorithms schedule their review, is becoming the default experience in several apps.

The deeper scientific question remains open: what is the optimal spacing function for a given material, a given learner, and a given retention target? Tabibian and colleagues framed this as a stochastic optimal control problem in 2019 [35]. Settles and Meeder at Duolingo approached it with the Half-Life Regression model in 2016 [36]. The FSRS project approaches it with gradient descent on massive review logs. Each approach produces slightly different schedules, but they all converge on the same core principle Ebbinghaus identified in 1885: memory is not permanent, forgetting follows a predictable curve, and review timed to that curve is the most efficient way to learn.

The gap between what Quizlet calls spaced repetition and what the scientific literature describes as spaced repetition is real, documented by Quizlet's own engineers, and grounded in different design goals. Understanding that gap does not require choosing sides. It requires understanding what each system is designed to do, and matching the tool to the task.

For more on how memory algorithms have evolved from SM-2 to modern schedulers, see FSRS vs SM-2: Spaced Repetition Algorithm. And for the neuroscience of how the brain tags information for long-term storage, see Sleep and Memory: The Brain That Stays Awake While You Sleep.

Conclusion

The story of spaced repetition spans 140 years, from Ebbinghaus sitting alone with nonsense syllables in Berlin to a Chinese undergraduate publishing an algorithm at a top machine learning conference. Along the way, a German science journalist built five cardboard boxes, a Polish student coded a formula in Turbo Pascal, an Australian programmer released an open-source app that medical students around the world now depend on, and a Silicon Valley startup built a flashcard platform used by hundreds of millions of people.

The science is not in dispute. Spacing works. Retrieval works. Sleep consolidation works. The meta-analyses are clear. The molecular mechanisms are increasingly well understood. The algorithmic implementations continue to improve.

What remains in dispute is the word "spaced repetition" itself. When Quizlet uses the phrase, it means adaptive within-session ordering based on predicted recall. When Anki or SuperMemo or FSRS-based apps use it, they mean cross-session interval scheduling with per-card state tracking. Both approaches are grounded in cognitive science. But they solve different problems for different users.

The student who types "Quizlet spaced repetition" into Google deserves to know the difference. Not to be told one tool is "better" and another is "worse." But to understand what each one actually does, so the choice matches the goal. That is what 140 years of memory science has earned: not a single answer, but a more honest question.