Spaced Repetition: The Science of Lasting Memory
How forgetting actually works, what the spacing research shows, and how to build a review system — intervals, card design, workload arithmetic, and limits.
By Velloreth
Most study advice is really advice about frequency dressed up as advice about technique. Read it again, highlight it, watch the lecture twice — all of it amounts to repeating exposure within a few hours of the same day, which is precisely the pattern the memory research says works least well for anything you need months later.
Spaced repetition is the correction: review the material at growing intervals, timed so each session happens shortly before you would have forgotten it. It is one of a small number of study techniques that has survived a century of experimental scrutiny, and it is also one of the most commonly misimplemented.
Forgetting is predictable
In 1885 Hermann Ebbinghaus measured his own retention for lists of nonsense syllables — deliberately meaningless material, to strip out prior knowledge — by tracking how much effort relearning took after various delays. He found a steep early drop that flattens over time: most of what you forget goes within the first stretch, and what survives the first few reviews decays much more slowly.
The curve’s shape is the whole insight. If forgetting is fast at first and slow later, then a review schedule should be dense at first and sparse later — the opposite of cramming everything the night before and never returning.
Each successful review flattens the curve further. Material recalled after a two-day gap decays more slowly than material recalled after two minutes, which is why the interval can grow each time: the memory is genuinely more durable, not just more familiar.
Ebbinghaus’s method had a real limitation worth knowing — nonsense syllables and self-reporting are not a classroom, and the original numbers do not transfer directly. What has transferred is the shape of the curve, replicated across more than a century of subsequent work.
What the spacing research establishes
Two findings matter most.
Spacing beats massing. A meta-analysis covering 184 articles, 317 experiments, and 839 assessments of distributed practice found spaced learning consistently better than massed learning, with an average benefit of roughly 15% across the studies reviewed — and the advantage growing as the gap between sessions increases, up to a point.
The optimal gap depends on how long you need the material. The same group ran a large follow-up: more than 1,350 people learned a set of facts, then reviewed after a gap ranging from minutes to months, then took a final test at delays up to a year. Recall was best at an intermediate gap for every retention interval — too short and the review was redundant, too long and forgetting had already done damage. The optimal gap came out at roughly 20% of the target retention interval for delays of a few weeks, falling to around 5% when the delay reached a year.
That gives you a usable rule of thumb:
| You want to remember for | Review gap |
|---|---|
| A test next week | 1–2 days |
| A semester (3 months) | ~2–4 weeks |
| Professional knowledge (1 year) | ~1 month |
| Long-term reference (years) | 1–3 months and growing |
The ratio matters less than the direction: longer retention target, longer spacing. A technique tuned for a quiz on Friday is not the technique for a certification in June.
Systems: Leitner, SM-2, and what software adds
Three implementations of the same idea:
Leitner boxes. Five physical boxes. New cards start in box 1 and are reviewed often; a successful review promotes the card to the next box, reviewed less often; a failure sends it back to box 1. Intervals are coarse because boxes are discrete, but the logic — promote on success, demote on failure — is exactly what software systems do with more resolution.
SM-2 (SuperMemo, 1987). The ancestor of most modern algorithms. Each card carries an ease factor starting at 2.5; a successful review multiplies the interval by that factor, a failure resets the card to a short interval. Repeated failures lower the ease, so hard cards shrink and easy cards stretch.
FSRS and similar modern schedulers. Instead of a fixed ease multiplier, these model retrievability directly — estimating the probability you currently remember the card and scheduling the review where that probability hits a target, typically around 90%. They learn parameters from your own review history. Anki supports FSRS, and for heavy users it produces shorter total study time for the same retention than the legacy algorithm.
The difference between these systems matters far less than two things the algorithm cannot provide: cards worth reviewing, and reviews that actually happen.
One caution about expanding intervals: a review of the literature found expanding schedules help, do nothing, or occasionally hurt compared with fixed spacing, depending on the study. Do not spend effort optimizing interval growth curves while your card quality and daily consistency are unsolved.
Making cards that encode something
This is where most spaced repetition systems fail. A card that asks you to recognize a sentence you copied from a textbook tests recognition of the sentence, not knowledge of the idea.
| Do | Don’t |
|---|---|
| One fact, one card | Bundle a five-step procedure into one card |
| Write the answer in your own words | Paste the source’s wording verbatim |
| Ask a question you could actually be asked in use | Ask trivia you only need for this deck |
| Include enough context to disambiguate | Strip all context so several answers fit |
| Use cloze deletion for a specific term in a sentence | Blank out half the sentence |
| Delete cards that no longer matter | Keep dead cards because you made them |
Writing the answer yourself is not busywork — it is the encoding step. Producing an explanation from scratch is also the same act as retrieval practice, which is the second high-utility technique in the research literature. The card then serves as the schedule, while the learning happened during its construction.
Practical card types, in rough order of power:
- Production questions — “Explain how a hash map handles collisions.” You must generate the answer.
- Cloze deletions — “TCP guarantees delivery through {{c1::sequence numbers and acknowledgements}}.” Context stays, the key element must be produced.
- Image or diagram prompts — “Label the parts of this circuit.” Good for spatial knowledge.
- Recognition cards — closest to multiple choice; weakest per the testing research, but still better than rereading.
Worked example: building a deck for one topic
Suppose you are learning HTTP status codes for a role that requires debugging APIs. A naive approach would copy a table of codes into cards: “200 = OK”, “404 = Not Found”, and so on. This produces recognition cards that test whether you can match a number to a label — a skill you will rarely need, because in practice you see the code and must decide what to do about it.
Better cards ask what you would actually be asked:
- “Your client receives a 429. What happened, and what does a well-behaved client do?” → “Rate limited. Back off, honor Retry-After if present, retry with exponential jitter.”
- “When does a 301 redirect matter for an API versus a 302?” → “301 is permanent — clients and caches update the stored URL. 302 is temporary — the original URL remains canonical.”
- “Your request returns 200 but the body contains an error object. Why might this happen, and what should you check?” → “Some APIs encode application-level failures in 200 responses. Check the body schema, not just the status line.”
Three cards, each encoding a decision rather than a definition. The first version would have produced ten cards you recognize; the second produces three you can use. Note the difference in what you wrote: the answer is in your words, grounded in a scenario, and specific enough that a wrong answer is diagnosable.
The same pattern applies to any domain: ask what you would need to produce under pressure, and write the card as that pressure.
Workload is arithmetic
Decks die from backlogs, not from bad algorithms. The load is predictable:
steady-state reviews/day ≈ new cards/day × average interval (days)
That is a model, not a measurement, but it captures the shape. Add 20 new cards a day and, once the average interval reaches 30 days, the queue approaches 600 reviews daily — an hour of work you did not anticipate when the deck had 40 cards. Add 5 a day and the same average interval yields about 150 reviews.
Consequences worth internalizing:
- Choose a sustainable new-card rate and keep it. A modest rate held for a year beats an aggressive rate abandoned in three weeks.
- Reviews dominate eventually. Early on you mostly meet new cards; later, most of your daily time goes to material you already learned. Budget for that shift.
- Leeches cost disproportionately. A card you fail repeatedly is usually a badly written card, not a fact you cannot learn. Rewrite it, split it, or suspend it.
- Skipping days is the real failure mode. The algorithm assumes reviews occur. Two skipped days convert into a wall of overdue cards that encourages more skipping.
When the queue grows: stop adding new cards first, work the backlog down, then resume at a lower rate. Deleting or suspending is a legitimate tool — coverage you abandoned entirely is worse than coverage you consciously trimmed.
Honest limits of the evidence
The spacing effect is robust; a few qualifications keep you from over-trusting it.
- Most foundational studies used artificial material. Ebbinghaus’s nonsense syllables, word lists, paired associates. The effect transfers to real learning, but the precise optimal intervals from laboratory studies are starting points, not constants — your own retention target and card difficulty shift them.
- Individual variation is real. Prior knowledge, sleep, stress, and the meaningfulness of the material all move the curve. Algorithms that assume one retrievability target for every card are models, not measurements of you.
- The comparison baseline matters. “Better than rereading” is a low bar that almost anything beats. The meaningful comparisons are against other active techniques — practice problems, self-explanation, interleaving — where spacing’s advantage is complementary rather than dominant.
- Implementation failure is common enough to be the main story. A well-designed algorithm with abandoned reviews loses to a crude Leitner box used daily. The effect size of the technique is conditional on the reviews actually happening.
What spaced repetition does not do
Be precise about its scope:
- It does not produce understanding. A card can encode “deadlock requires mutual exclusion, hold and wait, no preemption, and circular wait” without you being able to recognize a deadlock in a real system. Concepts need explanation, examples, and problems; cards retain the components those depend on.
- It does not teach discrimination between problem types. Choosing which technique applies is trained by interleaved practice — mixing problem types in one session — not by isolated fact cards.
- It does not replace practice for procedures. Typing, instrument handling, debugging, derivation: these need repetition with feedback, which looks more like deliberate practice than flashcards.
- Recognition is not recall. If you can only answer when the options are in front of you, the card format is too weak.
The productive framing: spaced repetition is the storage and retrieval layer of a study system. Encoding happens when you first wrestle with the material, understanding comes from explanation and worked problems, and scheduling comes from the algorithm. Each does a different job.
A workable routine
- Learn the material once, properly — read, solve, explain it aloud. Do not make cards directly from unread text.
- Write a small number of cards while it is fresh, in your own words, one idea each.
- Review daily, at a fixed time, before adding anything new. Reviews first, always.
- Cap the session (20–30 minutes) and reduce the new-card rate when it overflows.
- Rewrite on failure. If you miss a card twice, the card is the suspect.
- Let intervals grow. Resist the urge to manually bump a card because you “should” know it by now; the scheduler is measuring something you are not.
Two habits make the system compound rather than churn: writing cards yourself instead of clipping, and keeping the daily review appointment even when the queue is small. The research says spacing works. The system only inherits that benefit if the reviews happen on schedule and the cards are worth answering.