Working-Memory Training and Intelligence: What Changed?

How has the evidence changed?

Recent evidence changes the answer from a yes-or-no debate to a narrower question about which working-memory practice transfers, and to what. The September 2026 review, “The asymmetry of working memory training transfer,” reports a small signal from updating-style training on outcomes outside working memory. Yet its direct comparison with span training is marginal, and its intelligence-related analysis is supplementary. An earlier review, “Can we enhance working memory? Bias and effectiveness in cognitive training studies,” found no pooled fluid-intelligence effect across computerized training studies in healthy adults. Together, these findings make transfer more plausible, but do not establish that training reliably improves intelligence-test performance or general intelligence. Three outcomes must stay separate: improvement on the practiced task, transfer to a different working-memory task, and transfer to a more distant measure such as an intelligence test. Evidence for one does not automatically demonstrate the next. The 2026 review searched studies through September 2024, despite appearing in September 2026, so its publication date is not the date through which evidence was collected. The update is cautious: ask whether the particular training and untrained outcome have been tested directly, rather than treating either a blanket rejection or an IQ-raising claim as settled.

Sources: The Asymmetry of Working Memory Training Transfer: A Systematic Review and Meta-Analysis; Can we enhance working memory? Bias and effectiveness in cognitive training studies

What exactly changed in the 2026 evidence?

The change is methodological: the September 2026 review separates two kinds of working-memory practice, compares transfer in both directions, and examines outcomes outside working memory. That sharper question produces a more conditional picture than asking whether “brain training” works as one category. Its results are a synthesis across studies, not the outcome of one large training experiment. Working memory means temporarily keeping information available while also manipulating it. In span tasks, a person retains and recalls or recognizes items after a short delay. In updating tasks, the person monitors changing information and replaces what is no longer current. These task families overlap, but they do not make identical demands. The distinction lets the review test whether transfer depends only on similarity between practice and test, or whether the direction and demands of training matter too. In “The Asymmetry of Working Memory Training Transfer: A Systematic Review and Meta-Analysis,” the primary analysis combined 55 independent studies, 208 effect sizes, and 3,492 participants. The authors used a three-level meta-analysis because many studies contributed multiple outcomes; the model accounts for variation among effects within a study and across studies. They identified another 302 effect sizes for outcomes outside working memory, analyzed separately. Those counts describe a body of evidence with many measurements, not 302 independent studies or participants. The review reports effects as Hedges’ g, a small-sample-corrected standardized contrast between training and control-group change. It expresses an average difference in units of pooled variability, not IQ points, a percentile, or a predicted score change for one person. That distinction matters because the review pools unlike outcomes and studies; its standardized estimate cannot be read as a conversion to an intelligence-test scale. The scope also sets a boundary on what “2026 evidence” means. The review included healthy participants, excluded clinical populations, and searched research published through September 30, 2024, despite appearing in September 2026. Its publication date does not mean it incorporates studies from 2025 or 2026. The results sharpen comparison between training paradigms, while the search cutoff and selected population limit how broadly readers should apply them.

Sources: The Asymmetry of Working Memory Training Transfer: A Systematic Review and Meta-Analysis

Why does updating training look different from span practice?

In the 2026 review, updating training transferred to untrained span tasks, while span training did not reliably transfer to updating tasks. The direct comparison between directions was only marginal. This pattern may explain transfer within working memory; it does not show that updating reliably raises intelligence-test performance. Working memory is the temporary maintenance and manipulation of information. A span task asks someone to hold material briefly and then recall it. An updating task requires monitoring changing information, replacing contents as new material arrives, and maintaining relevant items. Both involve maintenance, but updating adds monitoring and replacement demands. “The asymmetry of working memory training transfer,” a 2026 systematic review and meta-analysis, included 55 studies, 208 effect sizes and 3,492 healthy participants. Updating-to-span transfer was Hedges’ g = 0.176 (95% confidence interval 0.104 to 0.249). Span-to-updating transfer was g = 0.048 (95% confidence interval −0.078 to 0.173), an interval that includes zero. Hedges’ g is a small-sample-corrected standardized group contrast, expressing the difference between training and control-group change relative to variation in performance. It is not an IQ-point scale, percentile, or forecast for one person. The apparent asymmetry needs care. A reliable estimate in one direction and an unreliable estimate in the other do not by themselves prove that the directions differ. The review’s direct comparison was marginal (p = .073). This is compatible with an updating advantage, but leaves uncertainty about its size. These are averages across studies, not guaranteed outcomes for every program. The authors propose a resource-demand account. Updating may recruit broader processes, including monitoring and manipulation, so practice could exercise resources also used by less demanding recall-focused span tasks. The reverse may be weaker if span practice does not exercise all operations required for continuous updating. This is an interpretation, not an established causal mechanism: the review compares results across studies rather than isolating each mental operation experimentally. The review required adaptive working-memory training and a control group, and included healthy samples. It reports that transfer direction varied with dose, participant age, and verbal versus nonverbal stimuli. These moderator findings are not personalized prescriptions for an ideal dose or material. Its literature search ended September 30, 2024; publication in 2026 does not mean it includes every study through 2026. This distinction sharpens the question but does not settle it. Combining updating and span programs could conceal different within-memory transfer patterns. Yet within-domain transfer cannot establish improvement on intelligence tests. The review analyzes non-working-memory outcomes separately, and its intelligence-related analysis is supplementary. Treat the proposed updating advantage as a hypothesis for testing on independent intelligence measures, not proof that practice changes a person’s IQ.

Sources: The Asymmetry of Working Memory Training Transfer: A Systematic Review and Meta-Analysis

Does the signal reach intelligence-test outcomes?

The September 2026 review reports a small, suggestive signal for intelligence-related outcomes after updating-style working-memory training. It does not establish a reliable benefit across intelligence tests: the main estimate combines several kinds of outcomes beyond working memory, while the intelligence-only result appears in a supplementary analysis. The distinction matters because evidence that a training program affects a broad category of outcomes is not the same as evidence that it improves a particular IQ test or a general-intelligence measure. In the review’s pooled analysis of non-working-memory outcomes, updating training had a Hedges’ g of 0.120, with a 95% confidence interval from 0.048 to 0.192 and p=.001. Span training had an estimate of 0.008, with a 95% confidence interval from −0.102 to 0.118 and p=.885. Hedges’ g is a small-sample-corrected standardized difference in group performance change relative to a control group. It expresses the contrast in standard-deviation units; it is not an IQ-point conversion. The estimated difference between the updating and span groups was marginal, F(1, 298)=3.174, p=.076. The review says the narrower supplementary analysis of intelligence-related measures showed a consistent pattern, but that statement alone does not establish the size, precision, or robustness of an intelligence-specific effect. (“The Asymmetry of Working Memory Training Transfer: A Systematic Review and Meta-Analysis.”) A common reading would be that updating “worked” because its estimate was statistically significant while span’s was not. That is not enough to show the training types truly differ. The relevant direct comparison asks whether their estimated effects differ from each other; here, that comparison was marginal rather than conventionally statistically significant. The non-working-memory result therefore supports a cautious possibility: the pooled evidence leans toward updating training having more transfer than span training, but uncertainty remains about the size and consistency of that difference. The supplementary intelligence pattern points in the same direction without converting the main pooled result into proof about every intelligence measure. Nor does a standardized group contrast predict what will happen to one person’s score. The review synthesizes differences across studies, measures, and participants. It does not provide an individual conversion into IQ points, establish lasting change, or show that any particular person would improve on a given test. Thus the newer evidence changes the question from whether all working-memory training transfers to whether certain training paradigms may produce small effects on particular untrained outcomes. For intelligence-test performance, that is a plausible route worth testing directly, not a dependable forecast or evidence that general intelligence has risen.

Sources: The Asymmetry of Working Memory Training Transfer: A Systematic Review and Meta-Analysis

Can a shared cognitive operation explain transfer?

A shared operation is a plausible reason for transfer, but a controlled 2023 study found that it was not sufficient by itself. In “Limits of near transfer: Content- and operation-specific effects of working memory training,” 171 young adults were randomly assigned across training groups and active or passive control conditions. After 12 adaptive sessions over two to three weeks, participants improved on the working-memory tasks they practiced. That improvement did not generalize to tasks measuring the same operation with different material. The result concerns transfer between working-memory tasks, not intelligence-test performance. The distinction is between what a task asks a person to do and the material on which they do it. Consider an illustration: during verbal updating practice, a person tracks changing spoken words, replacing an old item when a new one meets a rule. A transfer task might ask them to update visual symbols instead. The content changes, while the broad operation—monitoring and replacing current contents—remains. The researchers varied both content domain and working-memory operation, including storage and processing or relational integration. Their finding cautions against assuming that preserving an operation guarantees transfer when material changes. It does not show that all tasks or programs behave alike. The comparison condition also matters. The study found considerable differences when training groups were compared with active versus passive controls. An active control gives participants another activity; a passive control does not. Those comparisons answer different questions because the control experience differs. A gain relative to one comparison is not a context-free training effect. Improvement on practiced tasks can be real while the inference about why it occurred, or whether it carries to new outcomes, depends on the comparison. The September 2026 meta-analysis, “The asymmetry of working memory training transfer,” offers a related but distinct test. Across 55 studies and 3,492 participants in its primary analysis, it found a small transfer estimate from updating training to span tasks, while the reverse direction was not reliable. The review proposes that updating may place greater cognitive demands on participants, but that account is not proof of a single mechanism. The 2023 experiment tested whether shared strategies were enough to explain near transfer; the later review compared transfer direction between training paradigms. Neither result licenses a leap from a plausible mechanism to improved IQ-test scores. Researchers must measure performance on independent, untrained intelligence outcomes. A story about overlapping mental operations can guide what to test; it cannot replace the test.

Sources: Limits of near transfer: Content- and operation-specific effects of working memory training; The Asymmetry of Working Memory Training Transfer: A Systematic Review and Meta-Analysis

What does the adult-focused 2024 review say?

A 2024 meta-analysis of computerized cognitive training in healthy adults found no measurable average improvement in fluid intelligence across pooled comparisons. This does not prove every program has exactly zero effect; it means the adult evidence summarized there did not support a general benefit. The review also found no detectable link between working-memory gains and fluid-intelligence change. That is a comparison point for the newer, narrower claim about updating-style training.

“Can we enhance working memory? Bias and effectiveness in cognitive training studies” combined 33 independent comparisons involving 2,729 participants for its fluid-intelligence analysis. The standardized mean difference (SMD) expresses an average group contrast in standard-deviation units, not IQ points. The estimate was 0.007 (95% confidence interval −0.105 to 0.118; p=.908), near zero, with moderate between-study heterogeneity (I²=44%). The interval is compatible with small effects in either direction, but the pooled analysis found no statistically detectable average benefit. A meta-regression also found no significant association between working-memory training effects and fluid-intelligence effects (coefficient 0.135, SE 0.228, p=.556). This is a comparison across studies, not evidence about whether one person’s memory gain predicts that person’s score change.

Scope helps explain why this review neither cancels nor confirms the September 2026 result. It examined computerized programs aimed at several cognitive functions, not only working memory or a direct comparison of adaptive span and updating. Participants were healthy young and middle-aged adults, generally ages 18–64. “The asymmetry of working memory training transfer” instead separates updating and span paradigms and reports a small signal for updating across non-working-memory outcomes. The questions overlap, but interventions and outcome groupings differ. These are not head-to-head replications: a broad pooled null can coexist with a possible narrower effect, which still needs confirmation on relevant intelligence measures.

The 2024 review tried to limit practice effects in its main working-memory analysis by requiring assessment tasks to differ from training tasks; it analyzed very similar, criterion-like tasks separately. Familiarity with a task format can raise its score without showing broader transfer. The authors found much larger working-memory effects on similar criterion tasks than on less similar assessments. They did not apply the same strict control to every other cognitive outcome, so the fluid-intelligence result should be interpreted within the review’s methods, not as a universal verdict.

For adults, the review keeps the general claim in check: pooled computerized training did not improve fluid intelligence on average, and stronger working-memory gains did not reliably accompany stronger fluid-intelligence effects. A proposed exception needs controlled adult studies that isolate the program and measure untrained intelligence outcomes directly. Neither a pooled standardized effect nor a practiced-task change provides an individual IQ forecast.

Sources: Can we enhance working memory? Bias and effectiveness in cognitive training studies; The Asymmetry of Working Memory Training Transfer: A Systematic Review and Meta-Analysis

Why did the treated-control evidence remain skeptical?

The 2016 synthesis found no convincing far-transfer improvement when working-memory training was compared with a treated control, even though nearer working-memory outcomes improved. “Working Memory Training Does Not Improve Performance on Measures of Intelligence or Other Measures of ‘Far Transfer’: Evidence From a Meta-Analytic Review” combined 87 publications and 145 experimental comparisons with pretest, posttest, and a control group. Outcomes included nonverbal and verbal ability, word decoding, reading comprehension, and arithmetic. Immediately after training, participants showed reliable improvement on intermediate verbal and visuospatial working-memory measures. For outcomes farther from the training tasks, the review found no convincing evidence of reliable improvement against treated controls. Across studies, working-memory improvement did not track far-transfer effects. Success on a trained or closely related task, then, could not stand in for evidence of carryover to broader abilities. A treated control receives another activity rather than simply waiting or receiving no intervention. When similar in schedule and attention, it can help account for study participation, expectations, and extra effort during later assessment. It asks a more demanding question than whether the training group improves from its own baseline: did the target program produce more change than a credible comparison activity? This design does not remove every possible bias. The comparison activity might itself affect performance, participants may know which program is intended to help, and outcome measures still matter. Without an adequate comparison, a score increase has several possible explanations, making it difficult to attribute change specifically to training. “Can we enhance working memory? Bias and effectiveness in cognitive training studies” connects this concern to measurement. Its meta-analysis focused on healthy young and middle-aged adults, examined possible bias, and separated assessment tasks similar to training from different tasks. Apparent working-memory effects were much larger when assessment tasks resembled training. The review found no fluid-intelligence improvement and no relation between working-memory gains and fluid-intelligence change; its authors identify task choice as a substantial potential source of bias. These reviews do not provide a direct replication: they included different training programs and questions. They reinforce why near gains, especially on similar measures, should not be treated as evidence of far transfer. The skeptical conclusion should be bounded rather than absolute. The 2016 synthesis pooled diverse programs and outcomes and preceded analyses separating updating-style training from span practice. It challenges a general claim that working-memory training reliably improves intelligence or other distant outcomes; it does not directly test every narrower training paradigm. A newer paradigm-specific signal can make that possibility worth investigating, but it does not turn the older pooled evidence into a positive result. The practical standard remains direct evidence: an appropriate control group and untrained intelligence outcomes must show the claimed transfer. A broad training claim requires more than a better score on the practiced task.

Sources: Working Memory Training Does Not Improve Performance on Measures of Intelligence or Other Measures of “Far Transfer”: Evidence From a Meta-Analytic Review; Can we enhance working memory? Bias and effectiveness in cognitive training studies

Does long practice settle the question?

No. Two years of working-memory practice improved practiced tasks and a factor summarizing performance across them, but researchers observed no transfer to measured fluid or crystallized intelligence. “Training working memory for two years—No evidence of transfer to intelligence” tests the suggestion that a larger dose will eventually make practice generalize. It weakens the dose-alone argument without settling every program or population. The longitudinal study followed students initially in ninth grade. One hundred twelve students practiced a varied set of working-memory tasks every two weeks for two years. A separate group of 113 students completed pretests and posttests without the training program. The researchers assessed factors for working-memory capacity, fluid intelligence, and crystallized intelligence. A latent factor combines measures to represent a broader ability. The reported result was substantial, reliable improvement on the practiced tasks and on their working-memory factor; neither intelligence factor showed a training-related transfer effect. The abstract gives no effect-size estimate for the intelligence outcomes. The outcome choice matters. Improvement on tasks used in practice is evidence of learning those tasks. Improvement on a broader working-memory factor extends beyond one item format, but remains within the trained ability family. Fluid and crystallized intelligence measures asked whether change reached more distant outcomes. Their null result challenges the step from “practice improves working-memory performance” to “practice raises general ability.” It does not show that no participant improved, or predict results from every program. Duration makes this evidence informative, not universal. Two years is a substantial opportunity for change, so the null intelligence findings weaken the claim that too little training time alone explains earlier failures to detect transfer. Yet participants were school students, not adults, and the comparison group did not receive a matched training activity. An active control would better separate training from attention and expectations. The study trained a heterogeneous task set; it did not isolate every training paradigm or test the updating-versus-span distinction examined in “The asymmetry of working memory training transfer: A systematic review and meta-analysis.” That 2026 review reports a small pooled updating-training signal for non-working-memory outcomes, while its direct comparison with span training is marginal. The findings answer different questions: the student study tests whether prolonged mixed practice transferred to measured intelligence factors in one school sample; the review compares paradigms across healthy samples and varied outcomes. Together, they show that practice time alone does not guarantee broader change. Longer practice may strengthen practiced skills, but an intelligence-test transfer claim still needs a measured, untrained intelligence outcome, a suitable comparison, and a population relevant to the claim.

Sources: Training working memory for two years—No evidence of transfer to intelligence; The Asymmetry of Working Memory Training Transfer: A Systematic Review and Meta-Analysis

What does a score change actually establish?

A higher raw result establishes better observed performance on that measure under those conditions. It does not alone establish a population-relative IQ change, a lasting increase in general intelligence, or a causal effect of training. The inference depends on what was measured and whether a comparison links the change to training. Raw performance means the answers or task output recorded, such as correct responses. A normed interpretation compares performance with a defined reference group using a conversion supported for that instrument and population. “More items correct this time” describes observed performance; “higher relative standing” requires a relevant comparison. A raw score cannot supply it by itself. A standardized mean difference, such as the SMD reported in “Can we enhance working memory? Bias and effectiveness in cognitive training studies,” expresses an average difference between groups in standard-deviation units. It compares different scales; it is not an IQ-point conversion, percentile, or personal prediction. That 2024 review of healthy adults reported small working-memory gains overall and larger gains when assessment tasks closely resembled training tasks; it found no pooled improvement in fluid intelligence and no relationship between working-memory gains and fluid-intelligence change. Improvement on a practiced task therefore supports a narrower conclusion than improvement on a distinct intelligence measure. A change on a fresh measure still needs context. Format familiarity or a more effective strategy may improve performance without showing broad ability change. A single score contains measurement uncertainty, so a small difference may reflect ordinary variation. Repeated testing brings assessment practice; a controlled study can separate it from training effects. The 2016 review “Working Memory Training Does Not Improve Performance on Measures of Intelligence or Other Measures of ‘Far Transfer’: Evidence From a Meta-Analytic Review” analyzed 87 publications and 145 comparisons. It found reliable gains on intermediate working-memory measures, but no convincing far-transfer improvement over treated controls across ability, reading, and arithmetic outcomes. These findings constrain broad claims but do not explain one person's score change. For low-stakes practice, keep the conclusion bounded. Test IQ Free offers 50 original reasoning questions and a provisional estimated IQ range from an unvalidated conversion with no validated adult reference population. The displayed band is neither a confidence interval nor an established population comparison. Retaking the quiz cannot validate the conversion. The optional paid report adds explanation and practice structure, not measurement validity. A score change can prompt you to inspect which tasks became easier; by itself, it is not evidence that general intelligence changed.

Sources: Can we enhance working memory? Bias and effectiveness in cognitive training studies; Working Memory Training Does Not Improve Performance on Measures of Intelligence or Other Measures of “Far Transfer”: Evidence From a Meta-Analytic Review

What should a reader do next?

Practice a specific operation when the goal is to improve at that kind of task. For a broader claim, ask for improvement on fresh, untrained intelligence measures under controlled conditions, then check whether gains persist. The 2026 review, “The asymmetry of working memory training transfer,” found a small pooled signal beyond working memory for updating-style practice, but the direct comparison with span training was marginal and the intelligence-only pattern supplementary. The adult-focused 2024 review, “Can we enhance working memory? Bias and effectiveness in cognitive training studies,” found no pooled fluid-intelligence effect across broader computerized-training evidence. Together, these results support testing a narrow hypothesis, not predicting an individual score increase. For practice, take Test IQ Free’s 50-question reasoning test and review its worked answers. Treat its displayed IQ range as provisional: the conversion is unvalidated and has no validated adult reference population, so the range is neither a normed comparison nor a confidence interval. The optional $9 report adds explanation and practice structure, not measurement validity. If a decision affects education, clinical care, or eligibility, seek qualified assessment and consider multiple sources of evidence. The verdict would change with replicated controlled adult studies showing gains on distinct, untrained intelligence measures and relevant follow-up. Until then, practice can serve a task-skill goal without serving as evidence that general intelligence has risen.

Sources: The Asymmetry of Working Memory Training Transfer: A Systematic Review and Meta-Analysis; Can we enhance working memory? Bias and effectiveness in cognitive training studies

Questions readers ask