While the “modal model of memory” is still widely taught and accepted as a general theory, an enormous amount of recent research has focused on how short-term memory enables higher cognitive processes like those involved in planning, goals, and executive functions. Yet this research has revealed surprisingly intricate links between short- and long-term memory. Increasingly, it appears that interactions among prefrontal areas (traditionally thought to be important for short-term memory) and medial temporal lobe areas (traditionally thought to be important for long-term memory) are important for understanding a wide variety of behaviors. For example, some claim the hippocampus may be important for encoding novel actions or goals, and others claim that the prefrontal cortex is responsible for encoding and retrieval from long-term memory.
But it would be easy to carry this logic too far: anyone who’s taken Psych 101 is familiar with the famous case of HM, and similar patients who tragically illustrate the clear dissociability of long and short-term memory. In a recent article, Burgess & Hitch address this apparent paradox from the perspective of computational models of memory, and suggest that the problem for understanding the integrated-yet-dissociable nature of memory may emerge from a failure to consider the integrative role of “context” across all forms of memory.
The authors begin by reviewing behavioral evidence for a distinction between short-term and long-term memory in normal human subjects. This is an important step, because brain damage itself (as in the famous case of HM) might actually cause a dissociation between memory systems that doesn’t normally exist. Burgess & Hitch review how phonological but not semantic similarity between words impairs their immediate recall, and how the opposite characterizes recall over longer time frames; that short-term memory is thought to be subserved by sustained active firing whereas long-term memory is thought to be subserved by more lasting changes to synaptic efficacy; and that repeated presentation of a string of digits, even when other digit strings intervenes, allows for the gradual long-term learning of that sequence.
Burgess & Hitch suggest that an obvious computational mechanism for remembering the serial order of items (“temporal chaining,” in which each item is associated with the next in the sequence) seems to be incompatible with the kinds of errors that humans make in such tasks.
These errors typically involve order reversals or selective impairment only on those items that sound similar while sparing intervening items. Instead, they suggest that a variety of memory models are converging on an alternative computational solution, involving the association of an abstract “context signal” with each item which slowly changes over time (although the exact mechanism/mathematical function is controversial).
There is some fascinating evidence in support of such “context signal” models. For example, Burgess & Hitch review one study in which subjects were asked to study and recall a list of length n, and then asked to study and recall a list of another length, say m. Subjects tend to make what are known as “instrusion” errors on such lists, where they will accidentally recall an item from the n-length list when they should be recalling items from the m-length list.
However, what is most interesting is the sequential characteristics of these intrusion errors: at any given point in the second list, the item most likely to intrude from the previous list has the same relative position within the first list as the current position in the second list. In other words, if the first list contains “f, y, l” and the second list contains “r, e, q, o, x, j, u”, you are most likely to incorrectly recall “f” during the first third, “y” in the middle third, and “l” in the final third of the second list. This finding is consistent with the claim that items are “tagged” according to their position relative to the starting- and ending-context of a list, rather than being temporally chained to one another directly.
Even more fascinating evidence can be used to inform and constrain these models, according to Burgess & Hitch. For example, subjects are better at recalling serially-organized items if they are presented in a rhythmic way, relative to when items are presented evenly spaced in time. Similarly, subjects are better at remembering the first items in a list (the primacy effect) and the last items in a list (the recency effect) relative to items in the middle. However, in rhythmically grouped lists, they show primacy and recency effects within each rhythmic group, as well as across the whole list! Even when presented in rhythmic groups, the same temporally-relative patterns of intrusion errors occur, in which subjects are most likely to confuse items which are similarly positioned relative to the starts and ends of two lists regardless of differences in their total length. Burgess & Hitch suggest that computational models can account for these findings by assuming the processing of temporal intervals is “multidimensional” and self-similar across those dimensions, such that some context signals are temporally nested within other context signals.
In contrast, there are very different patterns of errors in longer-term memories. For example, there does appear to be some form of temporal chaining in long-term memory, such that items are better recalled in a forward order than a reverse order. This has been accounted for by a “temporal context model” in which the items themselves form the context that associates one item to the next (although in principle this seems very similar to the temporal chaining models derided above). Burgess & Hitch imply that this is the crucial difference between short and long-term memory: whereas short-term memory encodes more abstract elements of context, including temporal intervals and the representation of multiple nested contexts, long-term memory appears to encode contexts as a function of previously encoded items.
Burgess & Hitch then quickly review numerous models of memory which purportedly account for these findings. Personally, I don’t find any of these models super-compelling, so I’m not going to review them here. I’m also unsure about the utility of the distinction between the types of “context” that are supposedly important for the two memory systems. While the level of “contextual abstraction” seems like one good way to characterize the difference between prefrontal and hippocampal representations, I’m not sure it’s the “crucial” distinction. Furthermore it seems strange to imply that short-term memory cannot also utilize a context signal based on items themselves.
Nonetheless, Burgess & Hitch’s article presents a variety of challenges to all those who seek to understand human memory at a mechanistic level. Computational models of memory have tended to focus on one system or the other, and this article highlights the complexities that emerge when such putatively dissociable systems are networked by the healthy human brain.