Exemplar models of word representations have remained ambivalent or impressionistic as to precisely what veridical auditory information is stored in individual word exemplars. Earlier models (Johnson, 1997b) suggest all perceived information was stored in memory, whereas more recent proposals (Pierrehumbert, 2002; Goldinger, 2007) suggest some degree of abstraction occurs in storing particular exemplars. Findings from the phonetic accommodation paradigm (Goldinger, 1998; Nielsen, 2011, etc.) suggest that the accumulation of new exemplars may drive the spread of sound change. At the same time, some theories of sound change suggest that perceptual biases serve as a starting point for change (Ohala, 1981, 1983). The current study investigates how perceptual biases, such as the predictability of coarticulation, can shape the contents of exemplars. The experimental results suggest that an expected phonetic alteration, such as f0 raising on vowels following voiceless consonants—a predictable coarticulatory effect—is more likely to undergo some degree of abstraction when stored in exemplar memory, whereas unexpected phonetic detail (e.g., f0 raising following voiced consonants) is more faithfully stored or maintained for longer in memory. These findings suggest perceptual biases that could shape pools of exemplars, leading to different expectations for conditioned versus unconditioned sound changes.