Abstract
The memory updating (MU) process is a core component of working memory (WM). To systematically examine the validity of two commonly used MU tasks as WM measures, the present meta-analysis (76 studies, total N = 16,184) synthesized results on the correlation between the two MU tasks and two criterion tasks: working memory capacity (WMC) and fluid intelligence (Gf). Results indicated a moderate correlation between running memory (RM) and WMC ($r = 0.42$), a weak correlation between n-back and WMC ($r = 0.23$), and moderate correlations between both RM ($r = 0.40$) and n-back ($r = 0.34$) and Gf. Subgroup analyses showed that memory load moderated the correlation between RM and WMC, and stimulus-onset asynchrony moderated the correlation between n-back and both WMC and Gf. The recollection and recognition nature of RM and n-back contributed to their different correlation with WMC, and the involvement of controlled attention in both tasks accounted for their association with Gf. The present meta-analysis indicated that RM is a more valid WM measure in behavioral studies on individual differences.
Introduction: The Validity of Memory Updating Tasks
Working memory (WM) serves as the "mental workspace" essential for storing and manipulating information during complex cognitive tasks. A critical sub-process of WM is Memory Updating (MU)—the ability to dynamically replace old, irrelevant information with new, incoming data. Researchers frequently employ two primary tasks to measure MU:
- Running Memory (RM): Participants view a continuous stream of items and must recall the last $n$ items in the correct order.
- n-back: Participants view a stream of items and must determine if the current item matches the one presented $n$ steps prior.
Despite their popularity, a significant debate persists regarding whether these tasks truly measure the same underlying Working Memory Capacity (WMC) as "gold standard" measures like the Operation Span task. This meta-analysis aims to resolve this discrepancy by evaluating the convergent validity of RM and n-back tasks against WMC and fluid intelligence ($Gf$).
Methodology
The researchers conducted a comprehensive meta-analysis following PRISMA guidelines, aggregating data from 76 peer-reviewed articles. The final dataset comprised a total sample size of $N = 16,184$ healthy adults, providing substantial statistical power to identify true effect sizes. Beyond simple correlations, the study utilized subgroup (moderator) analyses to determine if task parameters—such as memory load, stimulus-onset asynchrony (SOA), stimulus modality (verbal vs. non-verbal), and response methods—influenced the observed validity.
Key Findings: The "Validity Gap"
The analysis revealed a clear distinction in how well these tasks reflect general working memory capacity. While both tasks are often used interchangeably in literature, their correlations with WMC differ significantly.
As shown in the results, Running Memory (RM) demonstrates a moderate correlation with WMC ($r = 0.42$), suggesting it is a robust proxy for capacity. In contrast, the n-back task shows a surprisingly weak correlation with WMC ($r = 0.23$), indicating it may capture different cognitive processes than traditional span tasks. Interestingly, both tasks correlate moderately with fluid intelligence ($Gf$), likely because both require controlled attention and the inhibition of "lures" (familiar but incorrect items), which are core components of general intelligence.
Moderator and Subgroup Analyses
The study further explored how specific task designs impact validity. For the RM task, using a "mixed" memory load (varying the number of items to be updated) significantly strengthened its correlation with WMC ($r = 0.48$). For the n-back task, the speed of presentation (SOA) was a critical factor; validity dropped when the presentation rate was slow (SOA > 2500 ms), suggesting that participants may switch to alternative strategies that do not rely on pure working memory when given excessive time.
Theoretical Implications: Recollection vs. Recognition
The theoretical distinction between the two tasks provides a likely explanation for these findings. RM is recollection-based, requiring participants to actively retrieve and generate items from scratch, which mirrors the demands of complex span tasks. Conversely, n-back is recognition-based, relying on a simple "yes/no" match. According to the Embedded-Processes Model, RM forces items into the "Focus of Attention" (FoA), whereas n-back can often be solved using "familiarity" cues from activated long-term memory, which does not tax core capacity as heavily.
Conclusion and Future Directions
The meta-analysis concludes that Running Memory is a more valid measure of working memory capacity for behavioral studies focusing on individual differences. While the n-back task remains ubiquitous, particularly in neuroimaging due to its recognition-based format, researchers should be cautious when using it as a standalone measure of WMC. Future research should continue to explore whether n-back's validity improves in specific domains or if its utility is primarily confined to measuring controlled attention rather than storage capacity.
Summary updated