What is learned from retrieving a memory that is not learned by studying the same information? In response to this question, I have proposed a new theory of retrieval-based learning in which I argue that retrieval strengthens the ability to completely activate all portions of a memory trace from an initial state of partial activation. In effect, retrieval serves to unitize the features of a memory, making the entire memory remain retrievable in the future when cue-related activation may be weaker. This theory, called the Primary and Convergent Retrieval (PCR) model, explains why practice tests produce both better long-term retention and faster retrieval than restudy of the same information. In this dissertation, I explore and test several predictions arising from the assumptions of the PCR model’s learning rule. In Experiment 1, I use evidence from retrieval latencies to demonstrate that even unsuccessful retrieval attempts produce learning. In Experiment 2, I demonstrate retrieval practice does not generalize between retrieval cues, which has important consequences for assumptions about what the features of memory representations may be, and retrieval routes through these features. And in Experiment 3, I show that when the same gradual unfolding of features that is assumed to allow learning during retrieval is deliberately engineered to occur during encoding, it produces the same types of retention and latency benefits produced by retrieval. These experiments further support the PCR model by confirming its prediction about what is learned from testing, and when this learning may be expected. Portions of this dissertation have appeared in previously published works; specifically, much of the general introduction and general discussion has appeared in Hopper & Huber (2018), and the entirety of Experiment 1 is also reported in Hopper & Huber (2019). Experiments 2 and 3 have not yet appeared in any published volume.