Distributed and drifting signals for working memory load in human cortex

Increasing working memory (WM) load incurs behavioral costs, and whether the neural constraints on behavioral costs are localized (i.e., emanating from the intraparietal sulcus) or distributed across cortex remains an active area of debate. In a pre-registered fMRI experiment, 12 humans (12 scanner-hours each) performed a visual WM task with varying memory load (0-4 items). We replicated a localized, load-dependent increase in univariate BOLD activity in parietal cortex. However, we also observed both systematic increases and decreases in univariate activity with load across the visual hierarchy. Importantly, multivariate activation patterns encoded WM load regardless of the direction of the univariate effect, arguing against a restricted locus of load signals in parietal cortex. Finally, we observed representational drift in activity patterns encoding memory load across scanning sessions. Our results suggest a distributed code for memory load that may be continually refined over time to support more efficient information storage.