An evolutionary model of memory


Essentially all animals have some kind of memory -- a way of storing information about experience. (A Nobel Prize for research on the molecular mechanism of memory was awarded to Eric Kandel for work with aplysia -- the sea slug.) If it's so widespread, it must be important. What's it good for? It's hard to imagine that evolution would select for something so an aging Physics Professor could sit on his porch nursing a beer and remembering his youth. Despite the fact that memory permits that, it must serve a more important purpose.

A model of Randy Buckner and Daniel Carroll at Harvard University (Trends in Cognitive Science 11:2, 2007) brings together decades of research on memory and puts memory in an evolutionary context. The main idea is this:

The point of memory for natural selection is not the past, but the future.

Animals remember experiences so as to know how to avoid predators, how to find food and mates. But since the future is not simply a direct replay of the past, memory has evolved to make prediction possible. Basically, little snippets of experience are stored with more explicit memories and combined to create an animal to make a pretty good guess as to "what happens next" (also called a "mental model"). This is called prospection and involves making decisions by "running" a mental model.

Buckner and Carroll suggest that a variety of useful memory tools have been built through through natural selection from this basic structure of memory:

  • Imagining future events (prospection)
  • Recalling past events (remembering)
  • Seeing things from someone else's perspective (empathy)
  • Navigation (moving through space)

They have carried out fMRI experiments that show approximately the same areas of the brain are involved in these four tasks. We are only particularly interested here in remembering, so I only show that part of their figure here.

One of the unexpected implications is:

Memory is not time stamped

For evolution, we need to know what's going to happen. Exactly when a previous experience occurred is irrelevant (though a memory may have associations that permit a reasonable inference of the time). We need to know the context that might change an appropriate action. We actually all know this. Almost everyone has had the experience of trying to reconstruct when something you remember well happened by finding associations with particular events that are tied to specific dates.

Another unexpected implication is:

Memory is reconstructed rather than replayed.

This means that memory is usually not "veridical" -- a true copy of events. It's usually recreated from small elements of memory associated with an event. This has lots of implications for the legal system. (How much should witnesses be trusted?) Lots of psychology experiments have shown that memories can be "planted" by suggestion and that the suggestee really feels that the non-existent event has happened.

This reconstruction is a big problem. It means that

We don't really remember the past, we predict it.

We can remember something that didn't happen or we can "cross-up" memories and link them in the wrong way. In this class you will often find that you "remember" an answer, only to learn that you have cross-linked two different memories to get a wrong result.

If we can't trust our memory, what can we trust? How else can we learn? In this class we will specifically use a variety of methods for "knowing-how-we-know" (epistemology); things like creating foothold principles that we can (almost always) trust, finding coherences and building a safety net of cross-links, and expressing things in a variety of ways.

Joe Redish 7/4/11



Article 230
Last Modified: February 8, 2019