Sequential Sampling as a Computational Principle of Episodic Memory

Abstract

Visual perception unfolds through sequential eye movements, yet computational models of episodic memory typically assume that event representations already exist. Here, we propose that sequential gaze sampling does not simply supply memory with visual information, but contributes to the construction and organisation of episodic representations. We introduce a computational model in which successive fixations are integrated into an evolving internal representation, subsequently stored in associative memory and later retrieved through partial reinstatement of the original sampling trajectory. Within to-be-remembered elements, successive samples support progressive individuation; between elements, more frequent transitions strengthen relational binding. As sampling proceeds, internal representations become progressively lower-dimensional while retaining information about where samples were taken and when they occurred in the sequence. Sampling history also remains embedded in the resulting representation: changing the order of otherwise identical samples alters both the representation itself and element-specific retrieval. At retrieval, reinstating more of the original sampling trajectory improves access to specific elements, while competition is strongest when memories share both semantic category and sampling history. Finally, an exploratory generative readout shows that regions sampled during encoding are reconstructed more faithfully than unsampled regions. Together, these findings suggest that sequential sampling provides a candidate computational process through which episodic representations are constructed, shaping their subsequent storage and retrieval.