TL;DR. Spacing is among the best-supported findings in learning science, but the evidence is lopsided. In the largest review (Cepeda and colleagues, 2006), 229 of 271 massed-versus-spaced comparisons (84.5%) tested memory within ten minutes, and one after a month or more. The one experiment that mapped review gaps against retention horizons (more than 1,350 people, trivia facts) found no universal best gap: 1 day for a 7-day horizon and 21 days for a 350-day one, roughly 30% of the horizon at 5 to 10 weeks and 6% at a year. Real skills have fewer trials: surgical residents given four weekly sessions retained a microsurgical skill better than residents given the same sessions in one day, although the groups did not differ significantly on the day training ended. The CEO move: give a retention budget only to skills where forgetting is expensive. The student move: accept that spaced, effortful practice feels worse while working better. In one experiment 78% of participants did better with spacing, and 78% rated massing as at least as good.
This is the scheduling companion to our ranking of 12 study techniques by research strength, where spaced practice sits second, and to the decay rates in how fast abilities fade without practice. It asks a narrower question: when the thing to retain is not a vocabulary list, what do you repeat, and when? For building the skill in the first place, see deliberate practice in the age of AI.
Why is almost all spaced-repetition advice about flashcards?
Because that is what the experiments used. In 1885 Hermann Ebbinghaus reported that 38 repetitions of a syllable series spread over three days did the work of 68 repetitions made on one day. His subject was himself, and when Murre and Dros replicated his forgetting curve in 2015 the design was unchanged: one person, about 70 hours of learning lists.
The field then scaled up the same task. The standard reference, Cepeda, Pashler, Vul, Wixted and Rohrer (2006), covers 839 assessments in 317 experiments from 184 articles, restricted by design to verbal memory. Across 271 direct comparisons, spaced study scored 47.3% correct against 36.7% for massed study, and only 12 comparisons showed no effect or a negative one. That robustness is why Dunlosky and colleagues (2013) gave distributed practice one of only two high-utility ratings among ten techniques. What the summary hides is when those studies tested memory.
Table 1. How long did the spacing literature wait before testing?
CEOtudent analysis of Cepeda et al. (2006), Table 1. Counts are as printed; the share column is our calculation.
| Time from last study session to test | Massed-versus-spaced comparisons | Share of 271 |
|---|---|---|
| 1 to 59 seconds | 105 | 38.7% |
| 1 minute to under 10 minutes | 124 | 45.8% |
| 10 minutes to under 1 day | 11 | 4.1% |
| 1 day | 15 | 5.5% |
| 2 to 7 days | 9 | 3.3% |
| 8 to 30 days | 6 | 2.2% |
| 31 days or more | 1 | 0.4% |
| All | 271 | 100% |
Source: Cepeda et al. (2006), Table 1.
Two rows hold 229 of the 271 comparisons: 84.5% of this evidence tested memory less than ten minutes after study. Seven comparisons (2.6%) waited more than a week, and one waited a month or longer. The review’s separate analysis of shorter versus longer gaps is similar: 433 of 539 performance differences (80.3%) come from retention intervals under two hours.
Spaced study scored higher in every one of these seven bins, so none of this weakens the effect. It weakens the schedules built on it. A plan to keep a skill alive for a year cannot be read off evidence that mostly stops at ten minutes, and the authors themselves called for fewer single-session studies and more research with meaningful retention intervals.
What did the experiment that mapped gaps against horizons find?
Two years later the same group ran that study. Cepeda, Vul, Rohrer, Wixted and Pashler (2008) taught more than 1,350 people (1,354 in the authors’ manuscript) 32 obscure trivia facts, gave one review after a gap of 0 to 105 days, and tested them 7, 35, 70 or 350 days later: 26 gap-by-delay combinations, with the amount of review held constant.
Table 2. Best tested review gap at each test delay
Delays, gaps, best gaps and recall gains as printed in Cepeda et al. (2008), authors’ accepted manuscript. The “share of delay” columns are CEOtudent calculations.
| Test delay | Gaps tested (days) | Best gap, recall | Share of delay | Best gap, recognition | Share of delay | Recall gain versus no gap |
|---|---|---|---|---|---|---|
| 7 days | 0, 1, 2, 7, 21, 105 | 1 day | 14.3% | 1 day | 14.3% | 10% |
| 35 days | 0, 1, 4, 7, 11, 21, 105 | 11 days | 31.4% | 7 days | 20.0% | 59% |
| 70 days | 0, 1, 7, 14, 21, 105 | 21 days | 30.0% | 7 days | 10.0% | 111% |
| 350 days | 0, 1, 7, 21, 35, 70, 105 | 21 days | 6.0% | 21 days | 6.0% | 77% |
Source: Cepeda et al. (2008), accepted manuscript.
Three results matter for scheduling a skill.
There is no best gap, only a best gap for a horizon. At every delay a longer gap first raised, then lowered, final performance. The authors’ conclusion: decide how long you need to remember before choosing a schedule.
The gap grows far more slowly than the horizon. The published abstract puts the best gap at about 20 to 40% of a 1-week delay and about 5 to 10% of a 1-year delay. A 50-fold longer delay moved the best tested gap from 1 day to 21 days.
Too short is worse than too long. Performance climbed steeply as the gap lengthened, then declined much more gradually.
Our earlier piece on fast upskilling protocols worked from the abstract and left the one-month and three-month rows blank. The full manuscript tests four delays and prints the parameters of the curve the authors fitted, so Table 3 fills them in.
Table 3. First review gap by retention horizon
CEOtudent analysis of Cepeda et al. (2008): extrapolations of a fact-learning study with one review session, not tested prescriptions for skills. Delays actually tested: 7, 35, 70 and 350 days. The 30-day, 90-day and 365-day rows borrow ratios from the nearest tested delay; 180 days was not tested. Every cell is recomputed in our calculation script.
| The skill must hold for | Ratio method | Authors’ fitted curve | Working rule (editorial) |
|---|---|---|---|
| 1 week (tested) | 1.4 to 2.8 days (abstract: 20 to 40%); best tested gap 1 day | 5.0 days (misses the data) | Next day or the day after |
| 1 month (nearest tested: 35 days) | 6.0 to 9.4 days (20.0 to 31.4%) | 12.2 days | About 1 week |
| 3 months (nearest tested: 70 days) | 9.0 to 27.0 days (10.0 to 30.0%) | 18.4 days | 2 to 3 weeks |
| 6 months (not tested) | None printed; an earlier lab study by the group (150 subjects) found a 1-month gap best | 21.4 days | 3 to 4 weeks |
| 1 year (nearest tested: 350 days) | 18.25 to 36.5 days (abstract: 5 to 10%); 21.9 days at the observed 6.0% | 23.7 days | 3 to 5 weeks |
The fitted-curve column deserves limited trust. Our implementation reproduces the paper’s worked example (a 23-day best gap, 7% of a 350-day delay) and matches the data’s winner at 35 and 350 days, but it ties 14 and 21 days at 70 days and picks 7 days where the data favoured 1. The authors call the curve reasonable, not uniquely accurate.
It also cannot recommend a gap longer than 30.6 days at any horizon, because that is where its forgetting-rate term bottoms out. That is a property of the function, not a finding. The same paper suggests reviews delayed by at least several months for retention over several years, and Bahrick and colleagues (1993), following 4 people who learned 300 foreign-word pairs over 9 years, found 13 sessions spaced 56 days apart comparable to 26 sessions spaced 14 days apart. Past one year, the one spacing study opened for this piece followed four people, so Table 3 stops there.
Two further checks. Kim and colleagues (2019) analysed workplace training data from 10,514 people and found the same pattern: the best spacing between retrieval events increased with the retention interval. And gaps need not keep growing. Across 29 studies, Latimier, Peyre and Ramus (2021) found spaced retrieval clearly better than massed (g = 0.74) but no significant advantage of expanding over uniform schedules (g = 0.034).
Does spacing work for skills that are not facts?
Yes, with thinner evidence. Donovan and Radosevich (1999) subtitled their meta-analysis “Now you see it, now you don’t”: across 63 studies and 112 effect sizes, spaced practice beat massed practice by a mean weighted effect size of 0.46, but the effect varied with task and interval, and less rigorous studies reported larger effects. The Cepeda review’s summary of it: the largest effects came from low-rigor studies of low-complexity tasks, and for skill tasks longer gaps produced smaller effects. An earlier tally cited there (Moss, 1996) found a benefit in over 80% of verbal and motor studies but only one third of intellectual-skill studies, with half showing no effect.
Table 4. Spacing evidence by skill type
CEOtudent editorial framework. Editorial grades: A = meta-analysis plus large studies with retention measured in months; B = randomized trials, but few, small or short; C = indirect or single small studies; D = no controlled spacing trial found. “Abstract only” = only the published abstract was opened.
| Skill type | Strongest studies opened | What was measured | Grade |
|---|---|---|---|
| Facts, terms, thresholds | Cepeda 2006 (271 comparisons); Cepeda 2008 (more than 1,350 people, up to 1 year); Kim 2019 (10,514 employees, abstract only) | Recall of facts; retention of workplace training | A |
| Procedures and checklists | Kerfoot 2014 (111 clinicians, abstract only): spaced question game against the same content posted online | Patients’ time to blood-pressure target: hazard ratio 1.043, called modest | B, thin: one trial, of knowing what to do, not of executing steps |
| Motor and technical skills | Moulton 2006 (38 surgical residents): 4 sessions in 1 day against 4 weekly. Spruit 2015 (abstract only): 65% of the spaced group proficient at the end of training against 21% of the massed group. Cecilio-Fernandes 2018 (abstract only): review of 11 randomized-trial articles | Retention at 1 month and transfer to a live model; retention at 2 weeks and 1 year | B: consistent direction, quality rated moderate |
| Concepts and problem solving | Rohrer 2020 (54 classes, 787 students): 61% against 38% one month later, d = 0.83. Rawson and Dunlosky 2011 (533 students, abstract only) | Unannounced mathematics test; recall of concepts after 1 to 4 months | B: strong trial, but of a package mixing spacing, interleaving and retrieval |
| Judgment and categorisation | Kerfoot 2010 (724 urology residents, abstract only): long-term score gains of 15.2% against 3.4%. Brunmair and Richter 2019 (59 studies, abstract only): g = 0.42 overall, 0.67 for paintings, -0.39 for words | Diagnosis from tissue images; classifying new examples | B for visual categories; C for calls with no agreed answer |
| Interpersonal and communication | None found | Not applicable | D, thin evidence: inference from the rows above |
Three details change how to read these studies.
The one-day course looks fine on the day. In the Moulton trial both groups received the same four sessions, and no outcome differed significantly straight after training. The gap opened a month later, on retention and transfer tests. Any course judged at the end of the day will flatter the compressed format.
Complexity cuts both ways. The older meta-analysis found the largest effects for simple tasks; in the laparoscopy study the spacing advantage was most pronounced for the more difficult tasks.
For technical skills, weekly is what was tested. Both surgical trials compared one day against weekly sessions and nothing in between, and the surgical review adds the caveat that matters: the optimal gap is unclear.
What is the card when the skill is not a fact?
A flashcard is three things: a prompt that forces retrieval before the answer, a reference to check against, and a date for the next attempt. Any skill can have all three. What changes is the unit. Three mechanisms carry over.
- Retrieve before feedback. In Roediger and Karpicke (2006), restudying won on a test five minutes later, lost on delayed tests, and still left restudiers more confident.
- Relearn to a clean rep on separate days. Rawson and Dunlosky (2011) prescribe recalling each concept correctly 3 times in the first session, then relearning it 3 times at widely spaced intervals.
- Mix what is easy to confuse. In the Rohrer trial no two consecutive problems shared a strategy, so learners had to choose the method, not just execute it. The authors add that new skills probably need some blocked practice first and that corrective feedback may be essential.
Table 5. The “What is the card?” framework
CEOtudent editorial framework, built on the studies in Table 4. Units and rules are editorial judgments, not tested protocols.
| Skill type | The card: the unit you repeat | How to space it | When not to use it |
|---|---|---|---|
| Facts, terms, thresholds | A question answered from memory before checking | First gap from Table 3; relearn to a correct answer each session | The fact is cheap to look up at the moment of use |
| Procedures and checklists | A timed cold run: every step without notes, then compare with the reference | First run within days; then about monthly if work rarely triggers it | The procedure changes faster than you review it |
| Tool workflows | A small real task from a blank file, tutorial closed | Separate days while learning; afterwards real use is the review | You already run it weekly |
| Motor and technical skills | A short, full-quality practice block with feedback | Separate days, not one long session; weekly worked in the surgical trials | No feedback is available, so errors get rehearsed |
| Concepts and problem solving | One problem from a shuffled set of old and new types, solved before reading the solution | Add earlier types to every new set | The topic is brand new: block a few first |
| Judgment and categorisation | An archived case judged in writing before the outcome is revealed | Mix confusable categories; bring missed cases back after the Table 3 gap | Categories defined by words, not examples (blocking did better) |
| Communication | A rehearsed opening or hard answer, spoken aloud and recorded | Short rehearsals on separate days before the event | As proof of anything: grade D evidence |
Each card also needs a rule for success and failure. The Kerfoot (2014) trial ran on three: a new question every 3 days, a repeat after 12 days if answered wrong and 24 days if right, retirement after two consecutive correct answers. A lapse halves the gap; two clean reps retire the item.
Where does an AI assistant fit, and where must it stay out?
For non-flashcard skills the bottleneck is card production, which is clerical work an assistant does well. Four jobs can be delegated: generate prompts and cases from your own reference material with the answers withheld; shuffle old and new types; check your finished attempt against the reference; schedule the next date under your rules.
The retrieval attempt cannot be delegated. Spacing works through effortful recall, and an assistant that supplies the answer turns a retrieval rep into a rereading rep. The rule is attempt first, in writing or aloud, before any answer appears. See what the research on cognitive offloading shows for why that boundary matters, and which AI tutors actually teach for which tools hold answers back. A general chatbot does not schedule reviews unless told to, so the schedule has to live in your plan.
How do you build a one-page skill retention plan?
Start with the CEO decision: which skills get a retention budget at all. A skill earns one only when failure at the moment of use is costly, real work does not exercise it often enough to act as review, and it cannot be looked up in time. Let everything else decay on purpose, the argument of how to choose what not to learn. Spacing also maintains only what was acquired; if the skill never reached competence, start with our time-to-competence data.
Table 6. Skill retention plan (one page per skill)
CEOtudent editorial framework.
| Field | What to write | Default if unsure |
|---|---|---|
| Skill and reason | What fails, and at what cost, if it is gone on the day it is needed | No clear cost, no budget |
| Horizon | Next likely real use, or how long it must hold | 3 months, then revise |
| The card | The unit from Table 5 | A cold attempt, never a re-read |
| First gap | From Table 3 | Next day for a week; a week for a month; 2 to 4 weeks beyond that |
| Expansion rule | What follows a clean rep | Hold or lengthen the gap, up to about a month |
| Lapse rule | What follows a failed rep | Relearn to one clean rep that day; halve the next gap |
| Retire rule | When it leaves the schedule | Two clean reps in a row, or real use becomes weekly |
| Review cap | Minutes per week across all skills | When the queue exceeds it, cut skills, do not stretch the cap |
Seven-question diagnostic (editorial; answer yes or no):
- Can you name the date each maintained skill must last until?
- Is every review a cold attempt before looking, not a re-read or re-watch?
- Does each attempt end with a check against a reference or a person?
- Did the first review fall on a later day, not later the same day?
- Is there a written rule for what follows a failed attempt?
- Are confusable items mixed, and brand-new ones blocked first?
- Do you keep a list of skills you have decided not to maintain?
Six or seven yes answers describe a system, three to five a habit, fewer than three cramming with good intentions.
What this does not mean
- Table 3 is not a tested schedule for skills. It extrapolates one study of trivia facts with a single review, and past the best gap performance declines, so longer is not always better.
- Spacing does not always win. Kornell and Bjork built a word puzzle in which spreading the clues made the problem nearly unsolvable; massed presentation won, .34 against .22, with 20 participants. Blocking beat interleaving for word-based material (g = -0.39).
- Your sense of what worked is not evidence. In Kornell and Bjork’s main experiment (120 participants studying paintings by 12 artists), 78% did better when the artists were interleaved, and 78% said massing was as good or better. Judge a schedule by a delayed cold attempt, the same lesson as tutorial hell.
- This is not a decay forecast or clinical guidance. Medical studies are cited as training research only.
Frequently asked questions
Is spaced repetition only for memorising facts?
No. Randomized trials show benefits for surgical technique, diagnosis from images and mathematics. The evidence is thinner than for facts, and for communication skills no controlled trial was found.
How long should the first review wait?
It depends on how long the skill must last: a day or two for a week, about a week for a month, two to five weeks for a quarter to a year. These extrapolate a fact-learning study.
Should each gap be longer than the last?
Not necessarily. A meta-analysis of 29 studies found no significant advantage of expanding over uniform schedules.
Is an app necessary?
No. A calendar, a reference to check against and a lapse rule are enough. An AI assistant can generate and schedule items, provided the recall attempt stays yours.
Does spacing help with complex skills?
The record is mixed. One meta-analysis found the largest effects for simple tasks; a laparoscopy trial found the advantage most pronounced on the harder ones. Practise on separate days and test after a delay.
Sources
- Ebbinghaus, H. (1885). Memory: A Contribution to Experimental Psychology, Chapter VIII. Translated by Ruger and Bussenius (1913).
- Murre, J. M., and Dros, J. (2015). Replication and analysis of Ebbinghaus’ forgetting curve. PLOS ONE, 10(7), e0120644.
- Cepeda, N. J., et al. (2006). Distributed practice in verbal recall tasks: A review and quantitative synthesis. Psychological Bulletin, 132(3), 354-380.
- Cepeda, N. J., et al. (2008). Spacing effects in learning: A temporal ridgeline of optimal retention. Psychological Science, 19(11), 1095-1102. Published abstract plus authors’ accepted manuscript (ERIC ED505660).
- Dunlosky, J., et al. (2013). Improving students’ learning with effective learning techniques. Psychological Science in the Public Interest, 14(1), 4-58. Abstract only.
- Donovan, J. J., and Radosevich, D. J. (1999). A meta-analytic review of the distribution of practice effect: Now you see it, now you don’t. Journal of Applied Psychology, 84(5), 795-805. Abstract only; task findings via source 3.
- Moulton, C. A., et al. (2006). Teaching surgical skills: What kind of practice makes perfect? A randomized, controlled trial. Annals of Surgery, 244(3), 400-409.
- Spruit, E. N., Band, G. P., and Hamming, J. F. (2015). Increasing efficiency of surgical training: Effects of spacing practice on skill acquisition and retention in laparoscopy training. Surgical Endoscopy, 29(8), 2235-2243. Abstract only.
- Cecilio-Fernandes, D., et al. (2018). Avoiding surgical skill decay: A systematic review on the spacing of training sessions. Journal of Surgical Education, 75(2), 471-480. Abstract only.
- Kerfoot, B. P., et al. (2010). Online spaced education generates transfer and improves long-term retention of diagnostic skills: A randomized controlled trial. Journal of the American College of Surgeons, 211(3), 331-337. Abstract only.
- Kerfoot, B. P., et al. (2014). An online spaced-education game among clinicians improves their patients’ time to blood pressure control: A randomized controlled trial. Circulation: Cardiovascular Quality and Outcomes, 7(3), 468-474. Abstract only.
- Kim, A. S. N., et al. (2019). The spacing effect stands up to big data. Behavior Research Methods, 51(4), 1485-1497. Abstract only.
- Rohrer, D., et al. (2020). A randomized controlled trial of interleaved mathematics practice. Journal of Educational Psychology, 112(1), 40-52. Full text, grantee submission (ERIC ED595322).
- Bahrick, H. P., et al. (1993). Maintenance of foreign language vocabulary and the spacing effect. Psychological Science, 4(5), 316-321. Abstract only.
- Roediger, H. L., and Karpicke, J. D. (2006). Test-enhanced learning: Taking memory tests improves long-term retention. Psychological Science, 17(3), 249-255. Abstract only.
- Rawson, K. A., and Dunlosky, J. (2011). Optimizing schedules of retrieval practice for durable and efficient learning: How much is enough? Journal of Experimental Psychology: General, 140(3), 283-302. Abstract only.
- Kornell, N., and Bjork, R. A. (2008). Learning concepts and categories: Is spacing the “enemy of induction”? Psychological Science, 19(6), 585-592.
- Brunmair, M., and Richter, T. (2019). Similarity matters: A meta-analysis of interleaved learning and its moderators. Psychological Bulletin, 145(11), 1029-1052. Abstract only.
- Latimier, A., Peyre, H., and Ramus, F. (2021). A meta-analytic review of the benefit of spacing out retrieval practice episodes on retention. Educational Psychology Review, 33(3), 959-987. Abstract only.
Tables 1 and 3 and the share columns of Table 2 are CEOtudent analyses of the cited sources; Tables 4, 5 and 6 and the diagnostic are CEOtudent editorial frameworks. Of the 19 sources, 12 were read as published abstracts and are marked; the other 7 were read in full. Deliberately not used: any claim that people forget a fixed percentage within 24 hours, the 10,000-hour figure, any universal ladder of review days, and effect sizes or percentages from full texts that were not opened (Donovan and Radosevich’s task-complexity tables, Roediger and Karpicke’s recall scores).
This content was compiled with the support of AI following in-depth research, then written and prepared for publication by the CEOtudent editorial team.
This post is also available in:
