Why is subsequent recall sometimes better for self-generated answers than for answers obtained from an external source (e.g., calculator)? In this study, we explore the relative contribution of 2 processes, recall attempts and self-computation, to this "generation effect" (i.e., enhanced answer recall relative to when problems are practiced with a calculator). Adults (N = 36) practiced unfamiliar alphabet arithmetic problems (A + 4 = ?; answer E), in 3 learning conditions: self-generating answers through recall or counting (self-generate learning), obtaining answers with a customized calculator (calculator-only learning), or using a calculator after first attempting answer recall (retrieve-else-calculator learning). Subsequently, participants were tested without a calculator. Retrieve-else-calculator learning was expected to produce an intermediate level of performance because it captures the benefits of recall attempts but excludes any benefits of the other self-generation process (mental computation). However, retrieve-else-calculator learning proved as effective as self-generation learning (latency, accuracy, and recall rates on test), and both led to a test performance that was superior to calculator-only learning. We suggest that mental generation processes that involve the retrieval of intermediate products may introduce retrieval interference that precludes a strong contribution to the generation effect. However, tedious mental computation may contribute to superior access indirectly, by inciting automatic recall attempts. This effect can be mimicked by imposing a delay prior to calculator use. Our results are compatible with the view that learning contexts that promote recall activities produce a customized type of knowledge to support cued-recall (Rickard & Bajic, 2006). (Contains 2 tables.)