Abstract
Dreams provide a unique window into human cognition, as their content may reflect individual and collective experiences. Here, we use language models to analyze 21,607 dream narratives, representing each report as a vector in a high-dimensional semantic space that enables continuous comparison without predefined categories. We ask whether different sources of information are organized within the same representational space and across multiple levels of semantic representation; we also tested how dream representational geometries evolve over time. We find that dreams produced by different individuals, as well as by individuals of different genders, occupy distinct regions of this space, particularly in abstract semantic representations. Similarity among descriptions of dreamers also predicts similarity in their dream content. Finally, we show how the semantic geometry of dream content evolves over time, with its trajectory shaped by the cumulative history of past reports. Together, this study provides evidence for the semantic structure of dream content and its evolution over time.