<p>Emotional events tend to be vividly remembered. While growing evidence suggests that emotions have their basis in brain-wide network interactions, it is unclear whether and how these whole-brain dynamics contribute to memory encoding. Here we combined functional MRI, graph theory, text analyses and pupillometry in a naturalistic context where participants recalled complex narratives in their own words. Across three independent datasets, emotionally arousing moments during narrative perception were associated with an integrated brain state characterized by increased cohesion across functional modules, which in turn predicted the fidelity of subsequent recall. Network integration mediated the influence of emotional arousal on recall fidelity, with consistent within- and between-network interactions supporting the mediation across datasets. Together, these results suggest that emotional arousal enhances memory encoding via strengthening functional integration across brain networks. Our findings advance a cross-level understanding of emotional memories that bridges large-scale brain network dynamics, affective states and ongoing cognition.</p>