Spaced Repetition

In 1885, the German psychologist Hermann Ebbinghaus memorised 2,300 nonsense syllables (zof, gex, tib) and tested himself at intervals. The plot of recall against time he produced is, to a first approximation, an exponential decay. Twenty minutes after learning, about 60% remained. After a day, about 30%. After a month, less than 20%. He did the experiments on himself, alone, for years.

His second discovery is the one the modern field is built on: every successful retrieval flattens the next decay curve. Memory is not a passive store that drains over time. Each act of remembering re-strengthens the trace, and the strengthening is larger when retrieval happens at the moment the trace is about to fail.

That is spaced repetition: schedule reviews at the expanding intervals where retrieval is hardest but still possible. The economics are sharp. With a small daily review budget (15–30 minutes for an active learner), it is possible to retain thousands of facts for years at roughly 90% accuracy.

At a glance

Each successful retrieval roughly doubles the next interval. The card moves from active rehearsal into long-term memory at near-zero ongoing cost.

The lineage

Five waves brought spaced repetition from a curiosity to an everyday tool:

Ebbinghaus (1885) — established the decay curve and the spacing effect.
Paul Pimsleur (1967) — applied to language learning, with concrete interval schedules ("graduated interval recall"). Pimsleur tapes for English speakers learning French still ship the same schedule.
Piotr Wozniak / SuperMemo SM-2 (1985–89) — the first algorithmic implementation. Each card carries an "easiness factor" that updates after every review; intervals expand or contract. SM-2 was published in 1989 and is the ancestor of every modern scheduler.
Anki (Damien Elmes, 2006) — open-source SM-2 implementation that won the consumer wave. Medical students and language learners adopted it heavily. Anki's .apkg deck format became a de-facto standard.
FSRS (Free Spaced Repetition Scheduler, Jarrett Ye, 2022) — replaces SM-2's hand-tuned heuristics with a memory model fit to anonymised review logs. Empirically outperforms SM-2 on workload-vs-retention trade-offs in published benchmarks. To verify: 2024 head-to-head FSRS-5 vs SM-2 numbers.

Each generation accepts the same core observation (the spacing effect) and improves the schedule.

Why it works (the modern biology)

Spaced repetition exploits two cellular phenomena:

Long-term potentiation (LTP) at hippocampal synapses requires NMDA-receptor activation patterns that are more efficient at spaced intervals than at massed ones. Bliss & Lømo (1973) characterised LTP; subsequent decades showed the spacing dependence.
Memory consolidation during sleep. Replay of hippocampal sequences during NREM sleep transfers labile memories into cortical long-term storage. Reviewing material just before failure increases the salience of that material for consolidation. Walker & Stickgold's 2000s work made this concrete.

The combination is what makes the schedule non-arbitrary: the optimal interval is the one at which the hippocampal-cortical replay has just transferred the memory but the trace is brittle. Review at that moment and the consolidation gets a second pass.

This is not just a learning algorithm; it is an exploitation of the biology.

What it does and does not solve

Spaced repetition is excellent for:

Vocabulary acquisition (any language).
Anatomy, taxonomy, drug-dose recall, classification systems — anything where the unit of knowledge is a card-sized fact.
Long-tail facts in medicine, law, religious scripture, and history that are referenced rarely but must be available cold.
Maintenance of dormant skills (a violinist's repertoire, a programmer's standard-library familiarity).

It is bad at:

Conceptual understanding. Cards trained to a definition do not transfer to derivation. Spaced repetition is a complement to active practice, not a substitute. See concept deliberate practice.
Skills with motor or perceptual components (driving, surgery, instrument performance). These require their own structured practice schedules.
Anything where the schema itself is unstable. If the underlying concept changes weekly, spaced repetition will re-strengthen yesterday's wrong version.

The literature is unambiguous on these limits. The popular online narrative often conflates "memorise via Anki" with "learn"; that conflation is the most common practitioner mistake.

Beyond the human user

The same algorithm shows up in machine-learning curricula. Curriculum learning (Bengio et al. 2009) introduces training examples in a difficulty-graded order. Replay buffers in reinforcement learning (Mnih et al., DQN 2015) preferentially replay surprising or under-learned transitions — a stochastic gradient version of Ebbinghaus's idea. Hippocampal replay is biology's version of the same trick.

The cross-pollination has become explicit. Recent ML papers (2023–24) explicitly cite human spaced-repetition schedulers as priors for curriculum design. To verify: a survey-level citation list.

Why this has to do with other realms

Spaced repetition is one of the few cognitive-science interventions with effect sizes that compete with concept deliberate practice in raw skill acquisition. It is operational concept second brain: the system remembers for you, but only by forcing you to actively remember.

It connects to concept compounding structurally: the gains from each review are small in isolation and large in aggregate over years. A medical student who reviews 100 cards per day for four years processes more than 150,000 retrievals — far beyond what unstructured study would produce.

It connects to concept attention economy negatively: the same recommendation-engine logic that exploits human attention can be inverted to exploit human memory, but the consumer market for tools that strengthen attention has been smaller and quieter than the market for tools that fragment it.

An open question

Is there an analogue of spaced repetition that works at the scale of an insight rather than a fact? A specific epiphany about concept emergence or concept quantum superposition is more than its restatement on a card, and current SRS tools cannot schedule the kind of half-remembered conceptual return that produces conceptual maturation. The cognitive-science problem is well-defined; the engineering problem is unsolved.

Key sources

Ebbinghaus, H. (1885), Über das Gedächtnis — the original monograph.
Cepeda et al. (2006), Psychological Bulletin, "Distributed practice in verbal recall tasks" — the canonical meta-analysis.
Wozniak, P. (1990), Memory Performance and Education and later SuperMemo papers — the SM-2 lineage.
Karpicke & Roediger (2008), Science, "The critical importance of retrieval for learning" — testing effect.
Ye, J. (2023), FSRS papers — modern fit-to-data schedulers.

Abhishek's take

I use this on the floor when a buyer has to remember why a red kurta failed in April, not just what the dashboard says today. A reminder at the next range review beats a training deck, because the question lands when the decision is about to repeat.

Abhishek