Humans continue to outperform modern AI systems in their ability to flexibly\nparse and understand complex visual scenes. Here, we present a novel module for\nvisual reasoning, the Guided Attention Model for (visual) Reasoning (GAMR),\nwhich instantiates an active vision theory -- positing that the brain solves\ncomplex visual reasoning problems dynamically -- via sequences of attention\nshifts to select and route task-relevant visual information into memory.\nExperiments on an array of visual reasoning tasks and datasets demonstrate\nGAMR's ability to learn visual routines in a robust and sample-efficient\nmanner. In addition, GAMR is shown to be capable of zero-shot generalization on\ncompletely novel reasoning tasks. Overall, our work provides computational\nsupport for cognitive theories that postulate the need for a critical interplay\nbetween attention and memory to dynamically maintain and manipulate\ntask-relevant visual information to solve complex visual reasoning tasks.\n