Caplan and colleagues have recently explained paired-associate learning and serial-order learning with a single-mechanism computational model by assuming differential degrees of isolation. Specifically, two items in a pair can be grouped together and associated to positional codes that are somewhat isolated from the rest of the items. In contrast, the degree of isolation among the studied items is lower in serial-order learning. One of the key predictions drawn from this theory is that any variables that help chunking of two adjacent items into a group should be beneficial to paired-associate learning, more than serial-order learning. To test this idea, the role of visual representations in memory for spoken verbal materials (i.e., imagery) was compared between two types of learning directly. Experiment 1 showed stronger effects of word concreteness and of concurrent presentation of irrelevant visual stimuli (dynamic visual noise: DVN) in paired-associate memory than in serial-order memory, consistent with the prediction. Experiment 2 revealed that the irrelevant visual stimuli effect was boosted when the participants had to actively maintain the information within working memory, rather than feed it to long-term memory for subsequent recall, due to cue overloading. This indicates that the sensory input from irrelevant visual stimuli can reach and affect visual representations of verbal items within working memory, and that this disruption can be attenuated when the information within working memory can be efficiently supported by long-term memory for subsequent recall.