Spaced Repetition

“We learn, in part, by forgetting.” —Robert A. Bjork

Spaced repetition is the practice of distributing study across increasing intervals of time rather than concentrating it into a single sitting. It is the load-bearing principle behind every serious flashcard system—see the companion Anki Explained essay for how the software puts it to work—and it is the most counter-intuitive of the three pillars of durable learning, because it deliberately makes each study session feel harder. These notes work from the empirical phenomenon (the spacing effect), through the theory of why it holds, to the mathematics and the scheduling algorithms that operationalise it.

The Spacing Effect#

The single most replicated result in the experimental study of learning is deceptively plain: for a fixed total amount of study, distributing that study across multiple sessions produces far better long-term retention than massing it into one. Ebbinghaus already noticed it in his founding monograph (Ebbinghaus, Hermann, 1885), observing that “with any considerable number of repetitions a suitable distribution of them over a space of time is decidedly more advantageous than the massing of them at a single time.”

The contrast holds time-on-task constant:

• Massed practice (“cramming”) — three hours in one sitting.
• Distributed practice — one hour on each of three separate days.

The massed learner almost always feels better prepared, and on an immediate test often scores slightly higher. But on any delayed test the spaced learner wins decisively. That gap between immediate confidence and delayed performance is the central pedagogical trap of the whole subject: the schedule that is best for tomorrow’s quiz is close to the worst for the final exam, and worse still for knowledge meant to last a lifetime.

The definitive quantitative summary is the meta-analysis of Cepeda, Nicholas J. and Pashler, Harold and Vul, Edward and Wixted, John T. and Rohrer, Doug (2006), pooling 254 studies and more than 14,000 participants: spaced practice yielded correct recall in 47% of cases against 37% for massed practice, and—crucially—the size of the advantage grew with the length of the retention interval.

Why Spacing Works#

The effect is an empirical regularity in search of a mechanism. The leading explanations are complementary rather than competing:

• Deficient processing. A second exposure that follows immediately on the first is processed only shallowly—the material is still in mind, so little work is done. Spacing forces fuller processing because the trace has partly decayed.
• Encoding variability. Studying a fact on different days binds it to different contexts, furnishing more independent retrieval routes than a single massed session yoked to one context.
• Study-phase retrieval. A spaced review is implicitly a test: to restudy an item you must first retrieve its fading trace, and that retrieval is itself memory-strengthening. On this view spaced repetition is not separate from active recall—it is active recall, scheduled.

The most influential modern framework is the New Theory of Disuse of Bjork, Robert A. and Bjork, Elizabeth L. (1992), which separates two properties of every memory:

• Storage strength — how durably entrenched a memory is. It only ever increases; well-stored knowledge is never truly lost.
• Retrieval strength — how accessible the memory is right now. It rises with study and cues but decays quickly with time and interference.

The counter-intuitive claim is that the two are inversely coupled at retrieval: the lower an item’s current retrieval strength, the more its storage strength is boosted by a successful recall. This is the principle Bjork named desirable difficulty—conditions that depress immediate performance (spacing, interleaving, testing rather than re-reading) tend to improve long-term retention. It yields the cardinal scheduling rule: review each item as late as possible while still recalling it successfully. Too late and the retrieval fails; too early and it is trivial and wasted. The sweet spot is the edge of forgetting.

The Mathematics of Forgetting#

To schedule reviews at that edge we need a model of decay. To a first approximation Ebbinghaus’ forgetting curve is exponential,

\[ R(t) = e^{-t/S}, \]

where \(R \in [0,1]\) is the probability of recall, \(t\) the time since the last review, and \(S\) the memory’s stability (the time constant of decay). The mechanism that makes spaced repetition work is that each successful, effortful retrieval multiplies \(S\): the curve resets to \(R \approx 1\) and then decays more slowly. If a well-timed review scales stability by a factor \(\alpha > 1\), then after \(n\) reviews

\[ S_n \approx \alpha^{\,n} S_0, \]

so the interval that holds retrievability at any fixed threshold grows geometrically—which is exactly why schedules expand: 1 day, 3 days, 1 week, 3 weeks, 2 months.

How long should the first gap be? It depends on how long you need to remember the material. Cepeda, Nicholas J. and Vul, Edward and Rohrer, Doug and Wixted, John T. and Pashler, Harold (2008) varied both the inter-study gap and the retention interval and traced the gap that maximised final recall—a “ridgeline” of optimal spacing. The qualitative law: the optimal gap grows with the retention interval, while the optimal ratio of gap to interval shrinks (roughly 20–40% for short horizons, falling toward 5–10% for very long ones). There is no single “correct” interval—only one proportional to how long you intend to keep the knowledge, which is precisely what an adaptive scheduler computes for you.

The Algorithms#

A spaced-repetition system (SRS) is an algorithm that, for each item, estimates its current retrievability and queues it for review when that estimate drops to a target threshold. Three designs trace the field’s history.

The Leitner System#

The first practical implementation needed no computer. In So lernt man lernen, the German science journalist Sebastian Leitner described a set of numbered boxes (Leitner, Sebastian, 1972). A correct answer promotes a card to the next box (a longer interval); a single mistake sends it back to Box 1. Each box is reviewed on its own cadence, so the scheme produces expanding intervals automatically and concentrates effort on the hard cards—all with cardboard and discipline. Its limit is granularity: a handful of discrete steps, and no notion of how wrong you were.

SuperMemo and SM-2#

In 1985 Piotr Woźniak began measuring his own forgetting curves; by 1987 he had built the first computerised SRS, SuperMemo (Woźniak, Piotr A. and Gorzelańczyk, Edward J., 1994). Its second-generation algorithm, SM-2, remains the most widely deployed spaced-repetition algorithm in history—it powered Anki for over a decade and is still its fallback. SM-2 attaches an ease factor \(\mathrm{EF}\) to each item and grades each review on a 0–5 quality scale \(q\):

def sm2(quality, repetitions, ef, interval):
    if quality >= 3:                       # correct response
        if repetitions == 0:
            interval = 1
        elif repetitions == 1:
            interval = 6
        else:
            interval = round(interval * ef)
        repetitions += 1
    else:                                  # incorrect: relearn from scratch
        repetitions = 0
        interval = 1

    ef = ef + (0.1 - (5 - quality) * (0.08 + (5 - quality) * 0.02))
    ef = max(1.3, ef)                      # floor: never schedule too aggressively

    return repetitions, ef, interval

Three design choices recur in every later system: intervals are multiplicative (the next is the previous times the ease factor, so \(S_n \approx \alpha^n S_0\) with \(\alpha = \mathrm{EF}\)); difficulty is learned per item (easy cards drift to a high ease and are seen rarely, hard cards recur often); and failure is expensive (a lapse resets the interval to one day). The weaknesses are just as instructive: resetting on every lapse discards hard-won stability, the ease mechanism is prone to the “ease hell” pathology where repeated lapses pin a card to the 1.3 floor, and the model has no explicit notion of retrievability—it cannot answer “what is the probability I recall this today?”, which is exactly what a scheduler should optimise.

FSRS#

The modern default in Anki is FSRS (the Free Spaced Repetition Scheduler), fitted on hundreds of millions of real reviews and descending from Woźniak’s later two-component model of memory (Woźniak, Piotr A. and Gorzelańczyk, Edward J., 1994). It represents each card by difficulty \(D\), stability \(S\), and retrievability \(R\). Two ideas separate it from SM-2. First, it predicts retrievability explicitly, so the scheduler can target a chosen retention (say 90%) and solve for the interval that hits it. Second, large-scale data show real forgetting follows a power law rather than a pure exponential, so FSRS uses

\[ R(t) = \left(1 + f\,\frac{t}{S}\right)^{c}, \qquad c < 0, \]

whose heavier tail captures the gentler decay of well-stabilised memories. Its parameters are fit to your own history, so it achieves the same retention as SM-2 with materially fewer reviews.

Using It Well#

• Trust the schedule. It already shows cards at the edge of forgetting; cramming ahead of an exam squanders the difficulty that does the work.
• Card quality is the bottleneck. No algorithm rescues a bad card. Prefer atomic prompts, demand production over recognition, and obey the minimum information principle.
• Hunt leeches. A few chronically-failed cards eat a disproportionate share of review time; rewrite, split, or delete them rather than grinding.
• Consistency beats intensity. Intervals assume daily review; ten minutes a day beats two hours a week.
• Understand first, memorise second. As Brown, Peter C. and Roediger III, Henry L. and McDaniel, Mark A. (2014) put it, the technique makes durable what you have already made sense of—it does not manufacture understanding.

The Wider Craft of Memory#

Spaced repetition is the empirically-grounded core of a much older craft. Long before the forgetting curve was measured, scholars cultivated the art of memory—the deliberate training of recall—and Francis Bacon set out its problem plainly four centuries ago.

Francis Bacon, The Advancement of Learning (1605)

Emblem reduceth conceits intellectual to images sensible, which strike the memory more; out of which axiom may be drawn larger and better precepts than as yet we have.

The classical answer to Bacon’s problem was the method of loci—vivid mental imagery pinned to imagined places—a technique dramatised in Joshua Foer’s account of the memory-championship circuit (Foer, Joshua, 2011). Such mnemonics make encoding unforgettable, but they say nothing about when to revisit a memory. Spaced repetition is the complement: it leaves the encoding to a well-made card and governs the schedule, turning a one-off feat of memory into durable, low-maintenance retention.

Retention, in turn, is only half of intellectual work—a fact you can recall on cue is inert until it is connected to others. That is the province of the note-taking tradition: Niklas Luhmann’s Zettelkasten, popularised for English readers by Sönke Ahrens (Ahrens, Sönke, 2017), weaves atomic notes into a growing web of cross-references. The two practices are natural partners. Spaced repetition keeps the atoms retrievable; a Zettelkasten builds the molecules of understanding from them.

References#

Ahrens, Sönke (2017). How to Take Smart Notes, CreateSpace Independent Publishing Platform.

Bjork, Robert A. and Bjork, Elizabeth L. (1992). A New Theory of Disuse and an Old Theory of Stimulus Fluctuation, Erlbaum.

Brown, Peter C. and Roediger III, Henry L. and McDaniel, Mark A. (2014). Make It Stick: The Science of Successful Learning, Belknap Press.

Cepeda, Nicholas J. and Pashler, Harold and Vul, Edward and Wixted, John T. and Rohrer, Doug (2006). Distributed Practice in Verbal Recall Tasks: A Review and Quantitative Synthesis, American Psychological Association.

Cepeda, Nicholas J. and Vul, Edward and Rohrer, Doug and Wixted, John T. and Pashler, Harold (2008). Spacing Effects in Learning: A Temporal Ridgeline of Optimal Retention, SAGE Publications.

Ebbinghaus, Hermann (1885). Memory: A Contribution to Experimental Psychology, Teachers College, Columbia University.

Foer, Joshua (2011). Moonwalking with Einstein: The Art and Science of Remembering Everything, Penguin Press.

Leitner, Sebastian (1972). So lernt man lernen: Der Weg zum Erfolg, Herder.

Woźniak, Piotr A. and Gorzelańczyk, Edward J. (1994). Optimization of Repetition Spacing in the Practice of Learning, Acta Neurobiologiae Experimentalis.