Drift-like dynamics and information flow across the cortical hierarchy during working memory

Working memory is supported by widespread and distributed brain regions spanning across the cortical hierarchy. However, how working memory content evolves and is transmitted across cortical regions remains largely unknown. Here, we investigated the flow of working memory information across the cortex using time-resolved fMRI decoding. Across multiple regions in visual and parietal cortex, we found that decoded working memory content drifted over time from the memoranda towards the later errors in memory reports, consistent with drift-like dynamics predicted by attractor models. These behaviorally predictive neural memory errors emerged earliest in higher-order dorsal visual area (V3AB) and intraparietal sulcus (IPS0), and then later in early visual cortex (V1), suggesting a propagation of mnemonic information originated in higher-level visual area. Inter-areal correlation analyses revealed that during memory maintenance, information flowed predominantly in a top-down manner--from higher to lower visual areas whereas during passive viewing feedforward dynamics prevailed. Together, our findings demonstrate that working memory maintenance involves temporally structured drift dynamics and feedback-dominated information flow across cortical hierarchy, providing a mechanistic link between neural population dynamics and the formation of memory errors.