Can You Train Working Memory? Part One: What the Effect Depends On

Abstract

The brain-training industry rests on a single claim: that practising a memory task makes you generally sharper. It is a testable claim, and it has been tested a great deal.

This series covers three papers on working memory training. Part One is a 2024 meta-analysis that separates the size of the effect from the way it is measured — and finds that most of the effect lives in the measurement.

Working memory improves by a standardised mean difference of 0.18. Restrict the analysis to assessment tasks resembling the training task and it becomes 1.15. Fluid intelligence does not move at all.

1.What Is Actually Being Claimed

Working memory is the system that holds information while you do something with it — the storage component and the executive component together. It is measured with complex span tasks, where you retain a list while solving something else, and with the n-back task, where you report whether the current item matches the one n items ago.

The field distinguishes near transfer from far transfer. Near transfer means the training improves similar tasks. Far transfer means it improves unrelated cognitive processes — most importantly fluid intelligence, the capacity to reason about novel problems.

Far transfer is the commercial claim. Nobody buys a subscription to become better at n-back. They buy it to become better at thinking.

The founding studies were published between 2002 and 2008 by Torkel Klingberg, Jesper Olesen and Susanne Jaeggi and their colleagues, reporting improvements in intelligence after short training regimes. As Jose Rodas and colleagues note in the meta-analysis discussed here, those results led directly to the founding of a company selling cognitive training — which is worth recording as the sequence of events that it was.

2.The Meta-Analysis

Jose Rodas, Afroditi Asimakopoulou and Ciara Greene, working between Universidad Espíritu Santo in Ecuador and University College Dublin, published this synthesis in Psychonomic Bulletin & Review in 2024. It covers computerised cognitive training in healthy adults only.

Two design decisions distinguish it. It is the first meta-analysis in this area to run a formal risk-of-bias assessment of its included studies using the Cochrane RoB 2 tool. And it deliberately excluded criterion tasks — assessments closely resembling the trained task — from its main analysis, then analysed them separately.

The main working memory analysis drew on 52 independent comparisons and 3,737 participants.

3.The Numbers

Training improved working memory by a standardised mean difference of 0.18, with a confidence interval from 0.093 to 0.268. That is small, and it is real: the interval excludes zero comfortably.

Restricted to criterion tasks — the assessments that most resemble what was trained — the effect was 1.15. Roughly six times larger.

Fluid intelligence, across 33 comparisons and 2,729 participants, did not move: a standardised mean difference of 0.007, p = 0.908, interval from −0.105 to 0.118.

That is worth reading carefully, because it is not the usual kind of null. It is not a study that failed to detect something; it is a tight interval centred almost exactly on zero, from a sample of nearly three thousand people. Whatever training does, it does not do this.

A meta-regression then tested whether people whose working memory improved most also gained most in fluid intelligence. They did not — the moderator was not significant. Even the responders did not transfer.

And a further set of meta-regressions found that neither the dose of training nor the type of programme predicted how well it worked. More training did not buy more improvement, which is difficult to reconcile with the idea that a capacity is being built.

4.Where the Effect Lives

The gap between 0.18 and 1.15 is the actual finding of this paper, and it has a simple explanation. If you train someone on n-back and then measure them with n-back, you have measured n-back.

The authors put the conclusion plainly: their analyses identified various potential sources of bias, with the most significant being the choice of assessment tasks.

This is not fraud and it is rarely even deliberate. A researcher who has built a training programme around a task reaches naturally for that task when measuring whether it worked. The result is a literature in which the headline number describes the practice effect rather than the capacity.

What survives the correction is genuine but modest: a small improvement in working memory measured on tasks that were not trained, and — from the paper's other analyses — improvements in some specific executive functions.

5.What This Is Not

It is not a publication bias artefact. The funnel plot was symmetric and a regression test found no asymmetry. The literature is not hiding its null results; the small effect is simply the effect.

This matters because publication bias is the usual explanation for a shrinking effect, and here it is ruled out. What remains is a measurement problem, not a reporting problem.

The risk-of-bias assessment did raise two concerns. Most studies did not report how their randomisation sequence was generated, and very few pre-registered their outcome measures and analysis plans — leaving open the possibility that what was reported was chosen after the data were seen. Excluding the high-risk studies did not change the results.

And the scope is healthy adults. Cognitive training in clinical populations — ADHD, mild cognitive impairment, rehabilitation after injury — is a separate literature with different aims, and nothing here bears on it.

Editorial Comment

MindHeaven® sells no cognitive training product, recommends none, and has no commercial interest in this question whatsoever.

We cover it because it is the cleanest worked example of the thing this library exists to do. Here is an effect that is real, small, and routinely reported at six times its size — not by inventing data, but by choosing the measurement that flatters it.

That manoeuvre is fully available in supplement research, and readers should watch for it in ours. When a trial reports an effect on the one outcome closest to the mechanism the product was designed around, the correct first question is whether any untrained, unrelated measure moved as well.

Part Two asks what happens when you simply tell participants the training will work.