Suddendorf & Corballis argue persuasively that the ability to foresee future “situations” is likely to depend on many different mental faculties, including memory of the past. But despite recognizing the complexity of the prediction process, the authors focus on a surprisingly narrow and problematic mechanism for it, namely, mental time travel.

As the authors formulate it, mental time travel essentially treats future as a version of the past: What one is able to “pre-live” about future events are those that one can relive about past events (target article, sect. 1, para. 1). The authors suggest that episodic memory helps “pre-live” future events, because it is this type memory that one needs for reliving the past. The key assumption here is that one can mentally create only those future events that one has specifically experienced in the past.

I contend that this is an unnecessarily narrow formulation of foresight, because one can obviously mentally create events that are sufficiently different from any that one has experienced before. The authors' formulation is also severely limiting because, if it were strictly true, it would mean that one would be able to foresee only those events that one has episodic memory of.

From the computational standpoint, it is clear that specific predictions about future events can, in principle, be made by using generic prior knowledge in a combinatorial fashion (see Glymour Reference Glymour2002). Information about the particularities of specific past events, such as that provided by episodic memory, is not needed. To cite a qualitative example, in order to foresee the possibility that I may be mugged if I walk through certain blocks of the city at night, I do not need the actual experience of having been mugged there at night. A general knowledge of risky time periods and risky neighborhoods is enough. This is because one can easily generalize and extrapolate, with arbitrary specificity and detail, from past experience. Thus, in the above example, one can not only foresee the possibility of being mugged, but also envisage the mugging event itself in arbitrary detail. Indeed, one can also vividly imagine events that one is certain never to have experienced in the past, such as a boulder rolling up a hill on its own. The point is that the authors' formulation of foresight ultimately amounts to placing patently untenable limitations on one's very ability to imagine.

Extending the authors' formulation of foresight to its logical limits, while perhaps not altogether fair to the authors, is nonetheless a useful exercise, because it reveals an instructive conundrum. To the extent that one can only foresee those future events that one has experienced in the past, and to the extent that events never repeat themselves exactly, one can never apply the memory of any past event to a future situation. Presumably, the authors would address this conundrum by allowing for some level of generalization and extrapolation, so that the future event does not have to be an exact replica of the past one. But that is precisely my point, too: Some degree of generalization and extrapolation is a prerequisite for predicting future events. But why limit it as severely and arbitrarily as the authors do?

The aforementioned logical exercise reveals another related, but more severe, computational limitation of the authors' formulation. Without the ability to extrapolate from generalities, the amount of particularities the brain would have to store would be subject to a combinatorial explosion. For every prediction of a future event, the memory of a corresponding past event would be needed. Conversely, what one can predict about the future will be limited by one's episodic memory. In the aforementioned mugging example, in order to foresee a mugging event, I would have to have the memory of having been previously mugged by the same person, and in the same city block, and so forth.

Again, the authors would presumably address this handicap by allowing some generalization across, and extrapolation from, past experiences. Doing so would, among other things, recognize that the various types of memory are not quite as distinct, and independent, from each other as one might think. That is, different forms of memory might interact with each other and with other mental faculties to help foresee the future. Although the authors allude to this possibility initially, they move away from it later, especially in rejecting several possible instances of foresight in nonhuman animals simply because they do not appear to involve episodic memory (target article, sect. 3).

Note that in terms of its amenability to generalization and extrapolation, episodic memory is the least suitable form of memory. That is, episodic memory by itself is a computational bottleneck. Therefore, other types of memory must play a major role, and mental time travel must play a correspondingly smaller role, in foresight.

The Bayesian estimation theory encapsulates the aforementioned general computational principles into a powerful and flexible framework for making predictions. Briefly, in this framework, prediction is a fairly straightforward extension of parameter estimation. The future value of a given parameter can be estimated by combining the relevant probabilistic information about the past and present values of the parameter (for more rigorous expositions, see Davidson & Wolpert Reference Davidson and Wolpert2005; Glymour Reference Glymour2002; Krauth Reference Krauth1983). Three features of the Bayesian framework are especially worth highlighting in this context. First, this framework is clearly biologically plausible. Second, in many cases, Bayesian prediction can be shown to be ideal. Third, the Bayesian framework is versatile, in that it can use all available relevant information, including different forms of memory, to arrive at a prediction. Thus, the Bayesian framework can utilize episodic memory, but is not dependent on it. In this sense, the Bayesian framework subsumes, and greatly extends, the authors' framework for foresight.

Of course, the Bayesian framework for prediction has its faults and limitations (see, e.g., Krauth Reference Krauth1983). But it represents, at a minimum, a substantive counterexample to the framework suggested by the authors.