Successful adoption of proven effective practice strategies such as distributed practice may contribute to much-needed improvement in mathematics performance. However, it is not yet fully understood if distributed practice is beneficial for long-term retention of complex procedural knowledge and, if so, for which initial practice performance level this spacing effect occurs. To investigate these two questions, we used a randomised between-subjects design (Practice Strategy: massed vs. distributed) with 61 primary school students' mathematical problem-solving performance as dependent variable. First, as hypothesised, we found a spacing effect on students' problem-solving performance. Second, again as hypothesised, we found that the magnitude of this spacing effect depended on their initial practice performance. Our findings imply that distributed practice leads to better long-term problem-solving performance than massed practice, but only for students with medium initial practice performance who have not yet completely mastered the task.