Proceedings of the Annual Meeting of the Cognitive Science Society

Recognition memory is strongly shaped by similarity structure, yet how item distinctiveness reorganizes the relationship between hits and false alarms is not well understood. Using distances in a multidimensional feature space to quantify item similarity, we examined item-level recognition performance for real-world objects. Low-distinctive items showed positive correlations between hit rates and false-alarm rates, reflecting variation in inter-item similarity as predicted by standard global-familiarity models. In contrast, highly distinctive items exhibited high hits but low false alarms, reducing or reversing the overall hit–false-alarm correlation. Split-half analyses showed that both hits and false alarms were reliable across participants, but their relationship depended on an item's position within the surrounding similarity space. The Hybrid-Similarity Exemplar model combining inter-item similarity with a distinctiveness-dependent self-match signal captured this reorganization. These findings show that memorability is reliable yet relational, emerging from similarity structure rather than intrinsic stimulus properties.