What makes some sentences more memorable than others? In this work, we treat the problem of recognizing previously seen sentences as a comparison between a target stimulus and noisy memory representations of previously presented stimuli. Building on past work in image and word memorability, we conduct a large-scale memorability experiment with 500 participants and 2,500 target sentences, eliciting variation in how accurately participants recognize repeated sentences. We predict the memorability of sentences from a) empirically established word-level memorability scores, and b) sentence-level distinctiveness and surprisal features that capture the compositional semantics of sentences. We find that the presence of individually memorable words is highly predictive of sentence memorability, but that sentence-level features also predict sentence memorability – especially in the absence of memorable words. This suggests that otherwise forgettable words can together create memorable compositional meanings that remain in memory and facilitate recognition.