Spaced Repetition & Long-Term Memory

Hermann Ebbinghaus was the first person to study memory scientifically. Between 1879 and 1885 he acted as his own experimental subject, memorising lists of nonsense syllables — meaningless consonant-vowel-consonant combinations like "DAX", "ZUB", "LEM" — and then testing his own recall at precise intervals. The nonsense was deliberate: Ebbinghaus wanted to strip away any advantage from prior knowledge or meaningful associations, measuring pure memory retention. In 1885 he published Über das Gedächtnis (On Memory), introducing the forgetting curve: a mathematical description of how quickly memory decays after learning.

The forgetting curve shows that memory loss is not gradual and linear — it is exponential. Ebbinghaus found that roughly 40–60% of new information is forgotten within one hour of learning, 60–70% within a day, and approximately 80% within a week — unless the information is reviewed. The curve then flattens: what survives one week tends to survive longer, but the critical window is the first 24 hours. Memory does not fade slowly and uniformly like a photograph left in sunlight; it collapses rapidly in the first hours and then stabilises. Murre and Dros (2015) replicated Ebbinghaus's original methodology 130 years later and confirmed the forgetting curve holds — the same exponential shape, the same dramatic early loss.

The forgetting curve also revealed something more useful: each time material is reviewed, the curve resets at a higher baseline and decays more slowly. After the first review, memory drops less steeply than after initial learning. After the second review, less steeply still. After three to four well-timed reviews, the forgetting curve becomes nearly flat — the material becomes what memory researchers call "durable." This is the core mechanism that spaced repetition exploits: not reviewing constantly, but reviewing at the right moments to intercept forgetting just before it completes.

The practical implication is uncomfortable: everything you studied last week, but have not reviewed since, has likely decayed by 60–80%. The knowledge is not entirely gone — there is usually residual memory that allows relearning to happen faster than initial learning — but it is not available for reliable retrieval. This is the foundational problem that spaced repetition solves. Not by asking you to study more, but by asking you to study differently: timed reviews at the optimal moment to intercept the forgetting curve before it drops to zero, repeatedly, until the memory becomes durable.

The spacing effect is one of the most replicated findings in cognitive psychology: distributing learning over multiple sessions produces stronger, more durable memory than the same total time spent in a single massed session. Massed practice — studying the same material for hours in one sitting — produces good immediate performance but rapid forgetting. Distributed practice — spreading study sessions over days or weeks — produces slower initial acquisition but dramatically better long-term retention. Ebbinghaus himself noticed this in 1885; the effect has been replicated in hundreds of studies since.

Cepeda et al. (2006) carried out the largest review of spacing research: 839 comparisons from 317 experiments in 184 articles on verbal recall. Spreading study out beat massing it in one session, and the longer you needed to remember something, the longer the best gap between sessions. A follow-up study (Cepeda et al., 2008) taught more than 1,350 people a set of facts and tested them up to a year later. With the same total study time, the best gap between sessions raised final recall by 64% compared with studying twice back to back. The advantage grew with the delay: recall was 10% higher at a one-week test, 59% at five weeks, 111% at ten weeks and 77% at a year.

The mechanism is related to the forgetting curve. When you cram, each repetition of the material occurs while the previous repetition is still fresh — retrieval is easy, requires little effort, and produces minimal memory strengthening. When you space practice, each repetition occurs after partial forgetting — retrieval requires real effort, and that effort strengthens the memory trace in a way that easy retrieval cannot. Bjork (1994) called this a "desirable difficulty": a learning condition that feels harder and performs worse in the short term but produces superior long-term retention. The discomfort of spaced practice — the feeling that you're starting from scratch at each session — is the mechanism working, not a sign of failure.

Cramming exploits a feature of memory that is useful for short-term survival but useless for education: working memory recency effects produce high performance immediately after study, giving the illusion that learning has occurred. But working memory fades within hours. The material was never consolidated into long-term memory because the spaced retrieval that forces consolidation never happened. Students who cram for Monday's exam will perform well on Monday. They will recall roughly 20% of the same material on Friday. The students who studied across the week will recall 70–80% on Friday — and the gap only grows over subsequent months. Spaced repetition is simply the deliberate implementation of distributed practice: a schedule that forces retrieval at the precise intervals that maximise consolidation.

Sebastian Leitner was a German science journalist who in 1972 published So lernt man lernen (How to Learn to Learn), describing what became the most influential analogue spaced repetition system ever devised: the Leitner box. The system is built from index cards and five physical compartments — originally cardboard dividers in a shoebox. Each card contains a single question on the front and answer on the back. The compartments represent increasing review intervals: cards in box 1 are reviewed daily, box 2 every two days, box 3 every week, box 4 every two weeks, box 5 every month.

The rules are simple: every new card starts in box 1. If you answer a card correctly, it is promoted to the next box. If you answer it incorrectly, it is demoted back to box 1 — regardless of which box it was in. The system is self-calibrating: cards that cause frequent errors cycle through box 1 many times, receiving frequent review. Cards you answer correctly every time move through the boxes and are reviewed only monthly. The result is an adaptive schedule where your most difficult material receives the most attention, and easy material is reviewed just often enough to maintain retention without wasting time.

The Leitner system is a direct implementation of the spacing effect applied to individual items. The five-box structure approximates the Ebbinghaus-optimal review schedule: each promotion doubles (roughly) the review interval, matching the flattening of the forgetting curve as memories consolidate. The demotion rule — back to box 1 on any error — is the active recall component: you cannot passively re-read a card and promote it. You must retrieve the answer. If retrieval fails, the card returns to the beginning of its journey, because a failed retrieval is evidence that the memory trace is not consolidated enough for the longer interval.

To build a Leitner box you need: a box or folder with five dividers, a stack of index cards, a pen, and a calendar. Write one question per card — front: question, back: answer. Put all new cards in box 1. Set a daily reminder to review box 1. Set calendar alerts for boxes 2–5 at their respective intervals. The critical discipline: review every day, even if only for 10 minutes. The system collapses if you skip sessions, because cards pile up in box 1 and the spacing schedule breaks down. This is the main limitation of the analogue system compared to software: it requires manual tracking and discipline that apps automate.

Piotr Wozniak was a Polish researcher who in the 1980s became obsessed with a question: what is the mathematically optimal interval to review a piece of information, given a target retention rate? The Leitner box uses fixed intervals that are the same for everyone. Wozniak realised that optimal intervals differ by item (some facts are harder than others) and by individual (some people have better retention for certain material). In 1987 he developed the first version of SuperMemo, a computer program implementing an adaptive spaced repetition schedule. In 1990 he published the SM-2 algorithm, which remains the mathematical foundation of Anki — the most widely used spaced repetition software in the world.

The SM-2 algorithm works as follows: each card has two variables — an interval (days until next review) and an ease factor (a multiplier representing how easy the card is for you personally). After each review, you rate your recall on a 0–5 scale. A rating of 5 (perfect recall) increases the ease factor; a rating of 1–2 (incorrect or very difficult) decreases it. The next interval is calculated as: interval × ease factor. A card with an ease factor of 2.5 (the default) reviewed today will be reviewed in 2.5 days, then 6.25 days, then 15.6 days — an exponentially increasing schedule. A card you repeatedly find difficult will have a low ease factor and short intervals; a card you always recall easily will have a high ease factor and long intervals. The system automatically allocates review time proportional to actual need.

Anki implements SM-2 with modern modifications. It is free, open-source, and runs on desktop and mobile. Its "daily reviews" feature shows you only the cards due today — you are never asked to review a card before its optimal window. After each review, you rate with Again (failed), Hard, Good, or Easy. Anki calculates the next interval accordingly. A typical user with 500 cards does 15–30 minutes of reviews daily once the system is established. The critical feature is consistency: Anki's algorithm is calibrated for daily use. Skipping days causes cards to pile up, and catching up produces a suboptimal compressed schedule. The discipline required is low — a few minutes daily — but that few minutes must be truly daily.

Wozniak's insight was that spaced repetition is fundamentally an optimisation problem: given a forgetting curve with known parameters, calculate the latest possible review date that still catches the memory before it falls below a target retention threshold (typically 90%). His research showed that for most learners, maintaining 90% retention on a large body of material requires surprisingly little daily review time — roughly 20 minutes per 1,000 cards in steady state. The alternative — no SRS, relying on massed re-reading before each exam — requires vastly more total time and produces far lower retention at any delay beyond the immediate exam period. The SRS investment pays compound interest: each review session reinforces previous reviews, and the intervals grow without bound in principle, so eventually you review important material once a year and still retain it.

Piotr Wozniak published "20 Rules for Formulating Knowledge in Learning" in 1999 — the most cited practical guide to flashcard design. The most important rule is rule 1: Do not learn if you do not understand. The second most important is rule 3: Build upon the basics. But for most learners building their first flashcard decks, the most practically impactful rule is rule 4: Minimum information principle — formulate cards so they contain the minimum amount of information necessary. A card that contains five facts on the back does not test five memories — it tests whether you can retrieve five facts simultaneously, which is a different (harder, less useful) skill. When you fail to recall all five, the algorithm cannot determine which of the five facts is the weak point. The card should be split into five atomic cards.

An atomic card contains exactly one piece of information. This sounds extreme but is correct. "What is the mitochondrion?" → "The powerhouse of the cell (produces ATP via oxidative phosphorylation)." These are actually two facts (the colloquial description, and the mechanism). They should be two cards: "What is the colloquial description of the mitochondrion?" and "What process does the mitochondrion use to produce ATP?" Each card tests one retrieval pathway. The algorithm can calibrate each independently. If you always recall the first and often fail the second, the second gets more frequent review while the first moves to long intervals.

Cloze deletion is often more effective than simple question-and-answer for factual material: rather than "What year did Ebbinghaus publish Über das Gedächtnis?" (answer: 1885), the cloze form is "Ebbinghaus published Über das Gedächtnis in [1885]." Cloze deletion provides more context on the front, reducing the chance of failing due to ambiguity rather than forgetting. For formulas, dates, and names, cloze cards typically produce higher accuracy and faster review than Q&A cards. For conceptual understanding, Q&A is superior — cloze provides too much context and can be answered by pattern-matching rather than genuine recall.

The most common mistake in building flashcard decks is copying content from notes verbatim. Copied text is rarely phrased to force retrieval — it is phrased to inform. "The spacing effect refers to the finding that distributed practice produces better retention than massed practice" is a note. "What effect does distributing practice over time have on retention compared to massed study?" is a card. The difference is the question structure: a card back is an answer to a question, not a statement. If you cannot easily convert a note into a question-answer pair, that is evidence that the note contains passive information — information that recognises well but does not retrieve well. Those items should be your first priority to convert into atomic, question-format cards.

A spaced repetition system has three components: a card creation workflow, a daily review ritual, and a triage process for managing weak items. The creation workflow determines what goes into the system. The daily review ritual determines whether the schedule is followed. The triage process determines whether difficult items are handled effectively or allowed to pile up. Most people who fail with spaced repetition fail at the creation workflow — they either add too much (copying notes verbatim, creating multi-fact cards, adding material they don't understand) or too little (only adding what feels important, missing gaps that only emerge later during review).

The most effective card creation workflow integrates with your existing reading or note-taking practice. After a reading session with Cornell notes, the cue-column questions are already potential flashcard fronts — the note entries are the backs. After a lecture, the summary box contains the key points that should become cards. After a problem set, the problem types that caused errors should become procedure cards. The discipline is to create cards immediately after the learning session, while the material is fresh enough to phrase the front as a genuine question (not a copied fact). Cards created a week after studying are typically vague and untestable — they were written from fading memory rather than from immediate understanding.

Daily review requires habit stacking: attaching the review session to an existing habit that happens at a fixed time. Reviewing immediately after breakfast, during a commute, or before starting work at a computer are common anchors. The session should be short — 10–20 minutes for a deck of 200–500 cards in steady state. Lally et al. (2010) showed that new habits take an average of 66 days to become automatic, with high variability. The first two months are the period of maximum dropout risk. During this period, reviewing every day is more important than reviewing perfectly — an imperfect five-minute session on a busy day is far better than skipping.

The triage process manages the inevitable cards that repeatedly fail. When a card is answered incorrectly three or more times, it is a candidate for redesign: the card may be too vague, too complex, testing something not fully understood, or lacking a memorable encoding. Apply Wozniak's rules: split the card if it contains multiple facts, rewrite the front as a clearer question, add a mnemonic image on the back if the fact is arbitrary. Delete cards for material you no longer need — a retired card is not wasted, it is appropriately triaged. Periodically, suspend cards for subjects you have completed and un-suspend before returning to those subjects. A well-maintained deck of 300 carefully designed atomic cards will produce better long-term retention than a poorly maintained deck of 3,000 copied notes.