TL;DR. The claim that trained skills collapse after a year of disuse traces back to one 1998 meta-analysis. Read its own retention-interval table row by row and two things stand out: 50% of its data points come from retention intervals of seven days or less, only 1.7% come from intervals beyond a year, and the column of effect sizes does not decline monotonically. A 2025 meta-analysis in Psychological Bulletin rebuilt the question with 1,344 effect sizes from 457 reports and reported half-lives instead: half of acquired gains lost after about 6.5 months on accuracy measures, 13 months on speed measures, 11 months on mixed measures. The single moderator that reliably slowed decay was having intervening opportunities to actually perform the skill. Complexity and real-world context helped too, both in the opposite direction from what most people assume.
A CEO does not maintain a factory by asking whether the machines still work. They ask how often each one needs servicing, and they schedule it before it fails. Your own capabilities deserve the same treatment, and for the first time there is a defensible number to schedule against. But you have to read the evidence carefully, because the most-quoted version of it is not what most people think it says.
The number everyone quotes, and where it actually comes from
If you have read anything about skill decay in the last decade, you have encountered a version of this: trained skills degrade badly with disuse, with effect sizes reaching d = -1.4 after a year without practice. That figure comes from Arthur, Bennett, Stanush and McNelly’s 1998 meta-analysis in Human Performance, which pooled 189 independent data points from 53 articles. It is a genuinely important paper and it has been cited for a quarter of a century.
Its Table 3 breaks the overall result into eight retention-interval bands. Almost nobody reads that table. We did, and we rebuilt it below with two columns the original does not print: what share of the total evidence base each band represents.
Original analysis: what the canonical skill-decay evidence base is actually made of
CEOtudent editorial framework. Data points, sample sizes and mean d values are read directly from Arthur et al. 1998 Table 3; the two share columns are our calculation. The data point column sums to 178 and the sample column sums to 8,719, matching the totals the paper reports, which confirms the reading.
| Retention interval | Data points | Share of data points | Sample size | Share of sample | Mean d |
|---|---|---|---|---|---|
| Less than 1 day | 3 | 1.7% | 45 | 0.5% | +0.02 |
| 1 to 7 days | 89 | 50.0% | 3,325 | 38.1% | -1.03 |
| 8 to 14 days | 13 | 7.3% | 680 | 7.8% | -1.54 |
| 15 to 28 days | 19 | 10.7% | 766 | 8.8% | -0.95 |
| 29 to 90 days | 33 | 18.5% | 2,656 | 30.5% | -0.65 |
| 91 to 180 days | 7 | 3.9% | 302 | 3.5% | -1.42 |
| 181 to 365 days | 11 | 6.2% | 670 | 7.7% | -1.04 |
| More than 365 days | 3 | 1.7% | 275 | 3.2% | -1.28 |
| Overall | 178 | 100% | 8,719 | 100% | -0.97 |
Three observations follow, and none of them are in the popular summary.
First, this is overwhelmingly a literature about the first week. Half of all data points sit in the 1-to-7-day band. Add the sub-24-hour band and 51.7% of the evidence base measures retention over a week or less. That is not a criticism of the authors, who were pooling what existed. It is a caution about what the pooled number means.
Second, the interval people actually care about is the thinnest. Only three data points, covering 275 participants, measure retention beyond one year. That is 1.7% of the evidence base carrying the weight of the paper’s most-quoted sentence.
Third, the column does not decline smoothly. Decay at 8 to 14 days (-1.54) is deeper than decay at more than 365 days (-1.28). The 29-to-90-day band (-0.65) is the shallowest of any band past the first day, shallower than the first week. Whatever the underlying process is, the row-by-row table is not the tidy exponential curve the concept is usually drawn as. It is worth noting too that the headline figure quoted from this paper, -1.4, is larger in magnitude than the more-than-365-days row in its own Table 3, and matches the 91-to-180-days row instead.
None of this makes skill decay unreal. It makes the canonical citation a weaker foundation for planning than its reputation suggests.
Two meta-analyses, one time window, twice the disagreement
Arthur and colleagues were not the last word. A second meta-analysis by Wang, Day, Kowollik, Schuelke and Hughes, published in 2013 in the volume Individual and Team Skill Decay, ran the same question and found decay that was, in their own framing, more modest: trivial immediately after acquisition (δ = -0.08) and rising to δ = -0.71 at a retention interval between 90 and 180 days.
Put that beside the Arthur table’s own 91-to-180-day row and you get a head-to-head nobody prints.
Original analysis: the same retention window, scored by three research teams
CEOtudent editorial framework. Each figure is reported by the cited source; the ratio and scale columns are our calculation.
| Source | Reports pooled | Effect sizes or data points | Estimate at roughly 90 to 180 days | Evidence base vs 1998 |
|---|---|---|---|---|
| Arthur, Bennett, Stanush and McNelly, 1998 | 53 | 189 | d = -1.42 | baseline |
| Wang, Day, Kowollik, Schuelke and Hughes, 2013 | not reported here | not reported here | δ = -0.71 | not comparable |
| Tatel and Ackerman, 2025 | 457 | 1,344 | reported as a slope, not a band | 8.6x reports, 7.1x effect sizes |
The two older estimates for the same window differ by exactly a factor of two. That is the honest state of the question before 2025: two credible teams, one time window, one estimate twice the other.
What the 2025 evidence actually says
Tatel and Ackerman’s meta-analysis, published in Psychological Bulletin, is the first study large enough to treat retention interval as a continuous variable rather than a set of bins. They pooled 1,344 effect sizes from 457 reports, 8.6 times the reports and 7.1 times the data points of the 1998 paper, and modelled decay as a rate.
Their headline results, all reported directly in the paper:
| Performance measure | Decay per month | 95% confidence interval | Acquisition gain (d) | Half of gains lost after |
|---|---|---|---|---|
| Accuracy-based | 0.08 SD | -0.10 to -0.05 | 1.49 (1.39 to 1.58) | about 6.5 months |
| Speed-based | 0.06 SD | -0.11 to -0.02 | 1.88 (1.68 to 2.09) | about 13 months |
| Mixed | 0.06 SD | -0.10 to -0.02 | 2.48 (2.19 to 2.77) | about 11 months |
For accuracy measures they also estimate that essentially all acquisition gains are gone after roughly 16 months. For speed and mixed measures they explicitly decline to estimate a total-loss point, because the projection would run past the range of their data. That refusal is a mark of a careful paper, and it is the kind of limit that gets stripped out when findings are summarised.
The most useful thing in the table is the ratio. 13 divided by 6.5 is exactly 2.00. The same underlying skill, left alone for the same number of months, appears to have decayed twice as fast if you test it for accuracy as if you test it for speed. Mixed measures sit at 1.69 times the accuracy half-life. “How fast does a skill fade” has no single answer, because the answer depends on what you point the stopwatch at.
A caution before you do the arithmetic yourself
There is a trap here that we walked into while checking the numbers, and it is worth flagging because it will catch anyone who tries to build a schedule from the monthly slope.
If you take the accuracy figures at face value, half of an acquisition gain of 1.49 is 0.745, and at 0.08 SD per month that implies 9.3 months to lose half. The paper reports 6.5. The gap is not a rounding difference, and it widens for the other two measures.
Original analysis: why the monthly slope is not a usable rule of thumb
CEOtudent editorial framework. The naive column is our own arithmetic on the paper’s published acquisition and slope figures; the reported column is the paper’s own model output. The gap is the point.
| Performance measure | Naive slope arithmetic | Paper’s reported half-life | Error if you use the shortcut |
|---|---|---|---|
| Accuracy-based | 9.3 months | 6.5 months | 43% too optimistic |
| Speed-based | 15.7 months | 13 months | 21% too optimistic |
| Mixed | 20.7 months | 11 months | 88% too optimistic |
The shortcut fails, and it fails in the dangerous direction every time: it tells you that you have longer than you do. The reason is that the paper’s meta-regression contains more than a straight line through months, so multiplying the slope by elapsed time does not reconstruct the model. Use the reported half-lives. Do not build a schedule out of the slope.
We are showing our failed derivation rather than hiding it, because the same shortcut is exactly what a summary of this paper would produce, and it would be wrong by up to 88%.
The three moderators that change the schedule
A decay rate tells you when to intervene. The moderator analyses tell you what intervention works. Three of Tatel and Ackerman’s findings matter for anyone planning their own maintenance, and two of them run against intuition.
Intervening performance opportunities slowed decay, but only for accuracy. Whether participants had any chance to actually perform the skill during the retention interval was a statistically significant moderator for accuracy-based measures (β = 0.07, p < 0.05). The same pattern appeared for mixed measures without reaching significance (β = 0.07), and was absent for speed (β = 0.01). This is the single most actionable result in the paper: occasional real use protects the accuracy of what you know, and does noticeably less for how fast you can do it.
Complex skills were retained better than simple ones. Component complexity was a significant moderator for accuracy and mixed measures (β = 0.03 and β = 0.07, both p < 0.05), and the pattern across all three complexity dimensions pointed the same way. High-complexity tasks held up better. Most people assume the opposite, and plan refreshers for the hardest thing they know when the fragile item is the simple procedure they rarely touch.
Real-world skills were retained better than laboratory ones. For accuracy measures, decay slopes were steeper for laboratory and artificial tasks than for every category of real-world task: medical and dental (β = 0.12), military and transportation (β = 0.16), sports (β = 0.18), and miscellaneous (β = 0.13). A skill learned and used in its real context decays more slowly than the same class of skill drilled in isolation.
Read together, these three point in one direction. Decay is not primarily a property of the skill. It is a property of how embedded the skill is in something you actually do.
Where this evidence does not apply
This matters more than any table above, and it is the part that gets lost when findings travel.
Tatel and Ackerman’s scope is procedural skills involving non-verbal motor components. Surgical technique, resuscitation, equipment operation, sports technique, flying. The half-lives of 6.5, 13 and 11 months are estimates for that class of skill. They are not measurements of how fast your Python fades, or your French, or your ability to read a balance sheet. Arthur and colleagues’ 1998 meta-analysis did include cognitive tasks, and found them more susceptible to loss than physical ones, but that finding sits on the thin, non-monotonic evidence base described above.
Anyone who tells you a language or a programming skill has a 6.5-month half-life is extending a motor-skill result past its data. We are not going to do that. What transfers is the structure of the finding, not the constant: decay is real, it is measurable in months rather than weeks for well-acquired skills, it is faster on accuracy than on speed, and intermittent real use is what slows it.
Both meta-regression models also carried significant residual heterogeneity (Q = 3,820.34 for accuracy, 898.41 for speed, 1,045.03 for mixed, all p < 0.01), which is the authors’ way of saying that a great deal of variation between studies remains unexplained even after the moderators. These are population averages, not predictions about you.
The maintenance schedule
Here is where the CEO half of the job takes over from the student half. The evidence gives you a decay rate and one reliable lever. Turning that into a schedule is a management decision, and we are labelling it as such.
CEOtudent Skill Maintenance Framework
This is a judgement built on top of the cited evidence, not a measured schedule. No study has tested these intervals. The reasoning behind each rule is stated so you can disagree with it.
Rule 1: service before the half-life, not after it. The reported half-lives are the point at which half the gain is already gone. A maintenance interval set at the half-life is a schedule for operating at 50% capability. For accuracy-critical procedural skills, the defensible planning interval is a fraction of 6.5 months, not 6.5 months. Quarterly is a reasonable read; it is not a measured one.
Rule 2: classify by what you are scored on, not by what the skill is called. The 2.00x gap between accuracy and speed half-lives means the same activity needs different servicing depending on which failure mode costs you. If being wrong is expensive, service on the accuracy clock. If being slow is expensive, you have roughly twice the runway.
Rule 3: a real performance opportunity beats a refresher session. This is the one rule with direct empirical support. Intervening chances to perform the skill significantly slowed accuracy decay. Arrange to genuinely use a skill occasionally rather than scheduling practice at it. Practising your negotiation technique is a refresher. Taking one small real negotiation is a performance opportunity, and the evidence favours the second.
Rule 4: audit the simple skills first. Complexity predicted better retention. The items most likely to have quietly degraded are the low-complexity procedures you consider beneath scheduling.
Rule 5: embedded beats isolated. Real-world tasks decayed more slowly than laboratory ones across every category tested. A skill you can attach to live work needs less deliberate maintenance than one you can only rehearse in a sandbox. Where you have the choice, embed it.
The student half of the discipline is accepting that this is a maintenance problem at all, rather than assuming that something once learned is owned. The CEO half is refusing to service everything at the same interval when the evidence says the intervals differ by a factor of two, and refusing to run the plant on an estimate you got by multiplying a slope.
If you are deciding what to invest in learning in the first place, the maintenance cost belongs in that decision. Our framework for which AI tools are worth learning deeply treats depth as a commitment rather than a preference, and the half-lives above are part of what that commitment costs. On how to build the skill before you maintain it, see our piece on deliberate practice in the age of AI. And if you are rebuilding a skill set later in a career, what the research says about learning after 40 covers the acquisition side of the same question.
Frequently asked questions
Is skill decay real, or is this just a measurement artefact?
It is real. Every meta-analysis reviewed here found meaningful loss with disuse, across four decades of primary studies and thousands of participants. The disagreement is about magnitude and timing, not existence.
Which number should I actually plan around?
The 2025 half-lives, if your skill is procedural and involves a motor component: roughly 6.5 months on accuracy, 13 on speed, 11 on mixed. For cognitive or knowledge skills there is no equally strong number available, and anyone offering one is extrapolating.
Why is the accuracy half-life so much shorter than the speed one?
The paper reports the difference rather than fully explaining it, and flags speed-accuracy tradeoffs as a methodological problem in this literature. One reading is that speed reflects a coarser, more robust motor pattern while accuracy depends on finer calibration that drifts sooner. Treat that as an interpretation, not a finding.
Does re-learning take as long as the original learning?
Not according to this literature, but Tatel and Ackerman specifically warn that many studies fail to isolate retention performance from relearning effects, which means the first measurement after a gap can already include some reacquisition. That is a caution about the evidence base, not an estimate of relearning speed.
Should I schedule refreshers or just use the skill occasionally?
Occasional genuine use, on the evidence available. Intervening performance opportunities were the moderator that reached significance for accuracy decay. Formal refresher training was not tested as a separate moderator in that analysis.
Do these numbers apply to a skill I never fully acquired?
No. Every figure above is expressed relative to a measured acquisition gain. If the skill was never brought to proficiency, there is no acquisition gain for the decay rate to erode, and none of this arithmetic describes your situation.
Sources
- Tatel and Ackerman, Procedural Skill Retention and Decay: A Meta-Analytic Review, Psychological Bulletin, 2025, DOI 10.1037/bul0000481. 1,344 effect sizes from 457 reports; meta-regression slopes and confidence intervals by performance measure; Table 6 acquisition effect sizes; the 6.5, 13 and 11 month half-life estimates and the 16 month total-loss estimate for accuracy; moderator results for intervening performance opportunities, component complexity and task type; residual heterogeneity Q statistics. Funded by the Army Research Institute for the Behavioral and Social Sciences.
- Arthur, Bennett, Stanush and McNelly, Factors That Influence Skill Decay and Retention: A Quantitative Review and Analysis, Human Performance, 1998, volume 11, issue 1, pages 57 to 101. Table 3 retention interval breakdown by data points, sample size and mean d; the overall meta-analytic estimate; the task-type and methodological moderator results; the statement on performance relative to pre-interval level after more than 365 days of nonuse.
- Wang, Day, Kowollik, Schuelke and Hughes, Factors Influencing Knowledge and Skill Decay After Training, in Individual and Team Skill Decay: The Science and Implications for Practice, edited by Arthur, Day, Bennett and Portrey, Routledge, 2013. Meta-analytic estimates of skill loss immediately after acquisition and at a retention interval between 90 and 180 days.
- Wang, Factors Influencing Knowledge and Skill Decay in Organizational Training: A Meta-Analysis, doctoral dissertation, University of Oklahoma, 2010. The underlying dissertation for the 2013 chapter estimates.
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:
