When having trouble retrieving an episodic event, such as where I put my keys, a common strategy is to run through the preceding events in order. For example, once I remember going out to get the mail, I'm more likely to remember hanging up my coat, and so on. This strategy works only because of a feature of episodic memory, which I'll refer to as temporally ordered retrieval. That is, the memory traces underlying episodic recall are organized such that the likelihood of retrieving information about an event at time t is significantly higher if someone is cued to retrieve information about an event at time t−1. Effects of this phenomenon have been well documented at the behavioral level, the most obvious being the asymmetrical contiguity effect in free recall (Healey & Kahana Reference Healey and Kahana2014), and the underlying temporal structure of memory representations is the subject of several computational hypotheses (Buzsáki Reference Buzsáki2005; Gallistel & King Reference Gallistel and King2009). Temporally ordered retrieval is a distinctive feature of episodic memory; semantic memory, by contrast, seems to be far more permissive about the types and directions of associations. For instance, although my semantic memories of the history of Turkey might be organized so that some temporally adjacent events are associated with one another, I might just as well have strong associations between stories about Anatolia, or facts related to the military.

Mahr & Csibra (M&C) provide a list of the distinctive features of episodic memory that form the basis for their argument that episodic memory serves a communicative function. These features delineate the general content of episodic memory and the manner in which the content is presented – but not how the content is retrieved or which contents are more likely to be retrieved. Although it would be unfair to expect the authors to discuss all of the distinctive features, I argue that the omission of temporally ordered retrieval, and of retrieval effects more generally, raises serious issues for their account.

I propose the following methodological principle for functional theories in an evolutionary context: If a distinctive feature of a system explains a significant number of the system's failures and successes, then this feature is likely relevant for understanding the function of the system. I'll now present a success of the episodic memory system that is best explained by appealing to temporally ordered retrieval. Then I'll discuss common memory failures that reflect problems with retrieval.

One way to get a handle on memory successes is to study memory experts. A technique of memory experts that has been documented as early as Roman times is the method of loci (MoL) (Cicero, Reference Cicero, Sutton and Rackham55 bce/1948). This technique is used by mnemonists such as Shereshevsky in Luria's famous case study (Luria Reference Luria and Solotaroff1987), but is also effective as an intervention in normal and clinical contexts (Dalgleish et al. Reference Dalgleish, Navrady, Bird, Hill, Dunn and Golden2013; Gross et al. Reference Gross, Brandt, Bandeen-Roche, Carlson, Stuart, Marsiske and Rebok2014). The therapeutic use of the MoL by Dalgleish et al. indicates that it is not only a neat party trick, but also an intervention that can increase fitness. In the MoL, a list or other kind of minimally structured set of items is learned by visualizing a well-known environment. Then the subject imagines herself walking through the environment and storing each item on the list in a different location. In short, the MoL takes semantic information and transposes it onto an episodic structure. Insofar as we are in a position to identify any memory successes, the MoL is an excellent candidate – and it relies on temporally ordered retrieval.

Now for failures. It's close to a consensus in research on all kinds of long-term memory that retrieval tends to be a functional bottleneck (Sweatt Reference Sweatt2010). One way of seeing this intuitively is to think of all the times you couldn't remember some fact that later came to you easily. That you can be cued into remembering in a different context indicates that the memory trace was there the whole time. Your failure was not a failure to encode or a failure to store the encoded trace for long enough, but a failure to retrieve the stored information. A clinical example is the selective retrieval of traumatic episodic memories in posttraumatic stress disorder (PTSD). Patients with PTSD experience retrieval of memories related to the traumatic event in an uncontrolled fashion in contexts in which these memories are not useful or relevant. PTSD is manifestly unhealthy for the patient and, therefore, a memory failure in the fitness sense.

In summary, temporally ordered retrieval is a feature of the episodic memory system that is critical for some significant memory successes and is implicated in other memory failures. Combined with the aforementioned methodological principle, we can conclude that a satisfactory theory of the function of episodic memory should involve an explanation of how the distinctive features of episodic retrieval contribute to that function, including but not limited to temporal ordering.

Can M&C accommodate temporally ordered retrieval into their framework? To do so, they would need to argue that it serves a communicative function. I'd be very interested to see how this could be accomplished, but I'll end by noting a possible obstacle. The example of the MoL suggests that episodic temporal ordering works closely with semantic retrieval – recall that the MoL records semantic information using the episodic system. To explain this relationship between the two memory systems and their respective content, M&C might have to extend their account to semantic memory function. In particular, I suspect that because the distinctively episodic mode of retrieval follows a standardized, inflexible rule for forming associations, it's more faithful than the mode used in semantic memory. In tasks where we're required to report unstructured information, especially in a fixed order, a faithful, standardized retrieval system is ideal. However, the more flexible semantic system for associations is more effective for other tasks. This suggests a function for the cooperation and division of labor between the two memory systems, which is unlikely to be only adaptive for communication.