Tests modify memory and can improve memory performance: Practice tests outperform additional study trials as a learning technique when the final memory test is difficult (e.g., delayed in time). This is referred to as the testing effect. Although existing theories propose single mechanisms to underlie this effect, the contributions of different cognitive processes are yet to be dissociated. Because most testing-effect accounts attribute the testing advantage to either encoding, maintenance, or retrieval processes, we propose a multinomial processing-tree model that disentangles the contributions of all three memory processes. By applying this model to testing-effect data, we show that (a) testing memory primarily creates maintenance benefits (i.e., resistance against forgetting) and that (b) the critical interaction of testing vs. study benefits with final-test delay is not driven by different retrieval strengths. Our results thus support maintenance accounts of the testing effect and are difficult to reconcile with retrieval-based explanations.