Abstract We propose a new theory for the benefits of recall practice based on intra-item learning. On this account, retrieval cues produce an initial memory state (termed ‘primary retrieval’). However, this state is incomplete and insufficient for overt recall of the item. A subsequent process, termed ‘convergent retrieval’, fills in any missing information through intra-item associations, allowing recall of the item. Because this occurs in a staged manner, directional learning occurs from the initially retrieved features to the subsequently retrieved features; in contrast, restudy produces less intra-item learning because restudy provides all features simultaneously. This account of the testing effect makes unique predictions regarding recall latencies. We confirmed these predictions in two experiments, examining recall latencies in free recall and cued recall. Specifically, for a final test taken immediately after a practice test that did not include accuracy feedback, restudy produced higher accuracy than test practice, but, at the same time, test practice produced faster recall than restudy. In other words, a comparison between accuracy and recall latencies suggests a process dissociation for the benefits of each type of practice. Alternative accounts of these effects were ruled out: (1) response order analyses of the free recall experiment ruled out cue-target associations; and (2) a cue-switching manipulation in the cued recall experiment (recall practice with cue A, final recall with cue B) ruled out context-target associations. According to the proposed theory, intra-item learning is narrow in one sense (i.e., unique to the cues used during practice), but robust in another sense (i.e., learning how to recall the item).