We extend an existing family of multinomial processing-tree (MPT) models for measuring cognitive processes in paired-associate learning to test-based learning and delayed memory testing. One paradigm that combines both aspects is the testing-effect paradigm: After studying a list of item pairs, participants restudy the same pairs (restudy group) or receive an initial cued-recall test without feedback (test group). At the end, participants complete a delayed final memory test, in which the test group typically outperforms the restudy group—a finding known as the testing effect (also retrieval-practice effect). Despite the wealth of research on this effect, its underlying mechanisms are still debated. Previous research has focused on single mechanisms and rarely aimed at dissociating memory processes that may contribute to it. To simplify the theoretical debate, the new family of MPT models disentangles the contributions of encoding, maintenance, and retrieval processes. This requires nothing more than (a) extending the final memory test to a free-recall test followed by a cued-recall test and (b) administering the final tests immediately versus after a delay. By applying various model versions to such data, it is shown that testing primarily creates maintenance benefits (e.g., resistance against forgetting), while the final retrievability of stored associations is not affected by testing versus restudying. The findings thus support maintenance accounts of the testing effect and challenge explanations assuming different accessibilities for initially tested versus restudied material at final recall. The new MPT model family can readily be applied to other test-based learning or delayed-testing paradigms.