Theories of word learning have proposed several tools that children and adults use to reduce the difficulty of the word learning problem. However, we propose that reducing difficulty may be detrimentaldifficulty may promote long-term word learning. In this study, we tested predictions of the retrieval effort hypothesis in a cross-situational word learning paradigm. Learners were presented with objects and labels in three conditions of learning (easy, medium, and difficult) and tested either immediately or one week later. Results revealed a counterintuitive pattern of performanceinitially, participants in the easy condition had the highest performance. However, after a one week delay, participants in the medium condition had the highest performance. Participants self-report of retrieval difficulty during learning is used to account for differences in performance over time. This work is discussed in terms of the implications for several fields of cognitive science: statistical learning, human memory, and language and cognitive development.