The encoding variability effect refers to a memory advantage that arises when the same materials are learned across multiple presentations that each emphasize different features of those materials. Prior work suggests that variability benefits are feature-specific: variable encoding of item features should aid item-information dependent tests (e.g., cued recall), and variable encoding of relational features should benefit relation-dependent tests (e.g., free recall). In our earlier studies we confirmed this pattern for variable encoding of item-specific features (via semantic orienting questions task) on a cued-recall test, but across two experiments, found surprising costs of encoding variable relational features (via story creation task) in free recall, compared to constant encoding condition. The present study aims to explain these recurring free-recall costs and to test whether any benefit of variability can be obtained for that test type. First, we ask whether covert retrieval practice – possible to utilize as a strategy in our constant-encoding condition but precluded in the variable condition – inflated performance in the constant condition. Second, we test whether combining encoding tasks within a single learning episode (introducing processing variability by mixing item-specific and relational tasks) produces superior free-recall performance compared with repeating either task twice in a variable manner. Results will clarify how feature and processing variability, together with retrieval demands, interact to produce either benefits or costs of encoding variability.