Abstract
Retrospective cues in working memory (WM) prioritize relevant information and guide attention toward target items to optimize behavior. Although the frontoparietal attention network is known to support this top-down attention mechanism, the broader cortical network behind this process remains elusive. Here, by combining EEG source localization with Bayesian non-negative matrix factorization, we identify nine WM task-related subnetworks, including two newly identified networks: the left parahippocampal-insula network (lPIN) and the right parahippocampal-insula network (rPIN). Dynamic network similarity (DNS) analysis reveals that the brain rapidly transitions into a 'silent' state following memory encoding. During retention, both retro- and neutral-cues reactivate memory representation; however, retro-cues elicit stronger reactivation, primarily involving the lPIN and rPIN. Furthermore, this reactivation correlates positively with behavior in the alpha band and negatively with lateralized alpha power. Additionally, the global network configuration during encoding shows significant behavioral relevance. In contrast, during the retrospective phase, a sparser network pattern with lPIN and rPIN as hub nodes significantly influences both behavior and alpha lateralization. These findings highlight the parahippocampal-insula network's crucial role in enhancing WM behavior through the flexible manipulation of target representations via retrospective attention, thereby supporting the attentional strengthening hypothesis of the retro-cue effect at the large-scale network level.