For lab meeting this week we're reading a testing effect review by Rowland (2014). It is a heroic effort. The main ideas are summed up well on p. 21, in the Conclusions section, where it is noted that a retrieval difficulty hypothesis is supported, whereas a transfer-appropriate processing account is not. Additional support was found for semantic elaboration accounts, like Mary's mediator hypothesis, but the bottom line is that currently, no one unifying account can explain all testing data.
Another interesting finding was that benefits of testing emerged immediately as well as a delay, which dovetails with some things Karpicke has been saying lately about the "crossover" in testing being an artifact of test difficulty and item selection effects.
The somewhat unsatisfying, but honest, conclusion is on p. 22, where it is stated that "the underlying mechanisms that produce the effect remain elusive," and that "the testing effect is likely to reflect multiple memory mechanisms." In other words, there's lots more work to be done.