Large-scale brain network reinstatement supports recognition memory.

Remembering past experiences requires the reinstatement of neural activity patterns, yet whether and how this process unfolds across large-scale brain networks remains unclear. Here, we developed an approach to construct dynamic large-scale brain networks from trial-wise activation maps in the Natural Scenes Dataset. We identified robust network-level item-specific representations at whole-brain, subsystem, and regional scales. Network-level reinstatement, particularly for connections involving higher-order visual and dorsal attention systems, was associated with condition-level recognition performance and predicted item-level recognition accuracy. Biologically, regional network representational patterns aligned with neuromodulatory receptor distributions and mitochondrial bioenergetic gradients. Moreover, network-level representations were associated with local activation-based representations and mediated their relationship with recognition accuracy. Nevertheless, the two representational forms exhibited distinct spatial topographies, contributed independently to recognition, and were anchored in different neurobiological architectures. Together, these findings uncover a previously unrecognized network-level representational mechanism that supports recognition memory, refining current models of distributed information coding in the human brain.