Implicit statistical learning, whereby regularities between stimuli are detected without conscious awareness, is importantfor language acquisition. This form of learning has often been assessed using measures that require conscious decisionmaking or explicit reflection (e.g., 2AFC tasks). We aimed to measure statistical learning more implicitly. We leveraged thefact that frequently co-occurring stimuli may be chunked into a single cognitive unit, reducing working memory demands.We developed an artificial grammar in which sequences contained pairs of stimuli which always co-occurred (chunks)and more variable between-chunk transitions. In a novel visual recall paradigm, participants were asked to rememberand recreate sequences of serially presented images. Recall of predictable sequences improved over the course of theexperiment. However, recall dropped to initial levels when participants were presented with random sequences containingno predictable chunks. This approach represents a valuable method to measure statistical learning implicitly, withoutrequiring conscious reflection.