Task goals dynamically reconfigure neural working memory representations

A hallmark of human intelligence is the ability to flexibly shift between tasks and modify our behaviour according to current goals and context. Working memory (WM) is thought to be critical for this cognitive flexibility. A key unanswered question is how neural representations of WM are configured to support different task goals. Thirty-four participants performed a shape WM task with different task rules in interleaved blocks: delayed match-to-sample (DMS) and delayed match-to-category (DMC), while electroencephalography (EEG) was collected. Our results show that stimuli from a circular shape space can be decoded from EEG across the memory delay. Neural stimulus patterns initially generalised across tasks but reconfigured to integrate task and category information and became task-dependent in the later delay. In the DMC task, neural patterns for shapes within a category became more similar than across categories, controlling for their physical distance. This category-dependent shift was associated with greater task-dependency in the neural stimulus representation. Together, our results reveal a dynamic reconfiguration of neural WM representations over the course of the delay period to support distinct task demands. Significance statementAs humans, we regularly move between tasks and change our behaviour according to the situation. This cognitive flexibility critically depends on our working memory, the ability to keep information in mind in the short term. It remains poorly understood how working memory configures mental representations to support different tasks. In this study, participants performed two tasks that probed working memory for visual details or object category respectively. Neural representations were initially similar irrespective of task but became more task-dependent over time in memory. The task-dependence in the neural representation was explained by a neural category bias that emerged during the memory delay. Our results demonstrate how neural representations integrate prior knowledge with stimulus representations in a task-dependent manner, supporting flexible cognition.