Remembering sentences not presented in lists

False memories are typically studied with lists of isolated words that converge on an unstudied critical item, as in the Deese–Roediger–McDermott (DRM) paradigm. Yet most real-world remembering concerns events expressed in richer linguistic forms. Here, we adapted the DRM to sentence materials and, using sentence embeddings, we arranged new sentence probes along a continuum of semantic similarity to each list's theme. Participants studied the DRM sentences under incidental encoding (Experiment 1), intentional encoding (Experiment 2), and imagination plus intentional encoding (Experiment 3). Across the three experiments, they also performed a classical version of the DRM with words presented following intentional encoding instructions. Across all experiments, semantic similarity robustly predicted false recognition for sentences, but this effect was strongest under incidental sentence encoding and weaker under the two intentional sentence encoding conditions. The correlation for participant-level sensitivity to semantic similarity between sentence- and word-based DRM was strongest when sentence encoding was incidental, and weaker when sentence encoding was intentional. By contrast, cross-task correlations for average critical-lure false alarms were comparatively stable across experiments and generally attributable to baseline endorsement of unrelated lures. These findings advance a quantitative account of false memory that extends beyond isolated words toward event-like language. Moreover, the findings indicate that sentence- and word-based DRM share a common semantic core and, more broadly, that what generalizes across paradigms is not false alarming per se, but reliance on semantic overlap as a basis for recognition.