Competition during working memory enhances long-term memory

Working memory (WM) continually interacts with the external environment, and this interaction may shape both immediate behavior and long-term memory. However, visual inputs that resemble what we hold in mind can sometimes facilitate and sometimes interfere with WM performance, posing a challenge for theoretical accounts of WM representational principles. Here, we test the possibility that this inconsistency may be explained by nonmonotonic interactions, like those observed between co-active representations in long-term memory. Specifically, interactions between perceptual and mnemonic content may follow a non-linear function of representational overlap, trading off between interference and facilitation at key similarity thresholds. To test this idea, we designed a task in which visual inputs appear during WM maintenance for naturalistic images. We systematically varied the visual similarity of this input to both behaviorally-relevant and -irrelevant WM items, using a deep convolutional neural network to objectively quantify relationships between stimuli. To assess how competition between WM and visual inputs shapes long-term memory (LTM), we later tested recognition for stimuli encountered in the WM task. Across two experiments, we found that moderate similarity between visual inputs and behaviorally relevant WM content yielded the best WM performance, whereas highly similar or dissimilar inputs impaired it. Thus, nonmonotonic interactions may explain why WM is sometimes robust and sometimes susceptible to sensory distraction. Strikingly, the same conditions that disrupted WM were associated with enhanced LTM, indicating that competition between internal and external representations can promote durable encoding. Together, these findings reveal how overlap between thought and perception shapes both momentary cognition and lasting memory.