Recognition memory asymmetries predicted by individual item memorability

According to current recognition memory theories, a major source of confusion when people judge which stimuli they remember is the contextual similarity of those stimuli to those from the studied list. However, these theories overlook the potential for confusions to arise from item-specific memorability properties not explained by list properties – namely, the likelihood that a stimulus will be correctly recognized if studied (“hittability”) or falsely recognized if not (“false-alarmability”) – independent of contextual similarity. Here, we introduce a memorability continuum model, which predicts that differences in the individual “hittability” and “false-alarmability” of arbitrary items A and B can often lead to asymmetric recognition performance in a forced-choice task such that recognition of target A over foil B is more accurate than recognition of target B over foil A. In a forced-choice experiment testing memory for real-world object images, the model successfully predicted specific item pair asymmetries. Using deep convolutional neural networks validated on human similarity judgments, we showed that memory-list context explains only a small fraction of variance in false-alarmability but not hittability. We also show that, while inter-item similarity explains a small percentage of variance in their forced-choice discriminability, ignoring their differential memorability properties as targets and foils fails to predict the asymmetries. Our results highlight the need to integrate item-level memorability into existing recognition memory theories.