Dyllan Simpson¹, Benjamin Johnson¹, Timothy Brady¹; ¹University of California San Diego

For decades, recognition memory models have successfully predicted human performance without incorporating any notion of intrinsic memorability, instead relying on similarity-based comparisons and decision criteria. Global matching models operate on the principle that memory decisions are a function of summed similarity to all stored items, plus a criterion for "old" responses. Despite this success, the notion of memorability as an intrinsic property of objects - independent of context and distinctiveness - has gained prominence in the field. In three behavioral experiments (N=352), we demonstrate how apparent memorability effects emerge from dataset composition and context. First, we manipulated the prevalence of supposedly memorable animate objects in the THINGS database, creating datasets that were either 90% or 10% animate images. When animate objects were overrepresented rather than rare, they became less memorable than inanimate objects (p<0.001), demonstrating that memorability reflects relative distinctiveness rather than intrinsic properties. If memorability depends on distinctiveness, it should vary with both encoding and test context. Our second experiment confirmed this by using the THINGS database but manipulating similarity during encoding: presenting "highly memorable" objects among similar items reversed established memorability patterns (p=0.019). Our third experiment manipulated test sequences while keeping encoding constant, showing that low-memorable objects tested among dissimilar items performed equivalently to high-memorable objects tested among similar items (p=0.43). Simulations using a global matching model (Shiffrin & Steyvers, 1997), captured all of these effects, as ‘memorability’ estimates decreased as objects’ feature distributions became more homogeneous. These results suggest that reliable memorability effects emerge from the relative distinctiveness of features in experimental datasets rather than being intrinsic to images. Memory performance reflects the interaction between an observer's representational similarity space and the distributional structure of both study and test contexts.