Memory Is Built, Not Played Back

The folk model of memory is a recording. Something happens, a copy is stored, and remembering is playing it back, imperfectly, perhaps, degraded by time, but essentially a matter of retrieval from a place where the original sits waiting.

Nothing about the way memory behaves fits that model.

Frederic Bartlett noticed it in the 1930s. He had English participants read a Native American folk tale, "The War of the Ghosts", and retell it after various delays. The retellings did not simply lose detail. They changed shape: unfamiliar elements were dropped, the supernatural was rationalised, canoes became boats, and the story drifted steadily toward something an English reader of the 1930s would find coherent. People were not reproducing the tale. They were rebuilding it from fragments plus expectation, and they were confident about the result.

Elizabeth Loftus turned this into an experimental programme with unsettling precision. In the best-known study, participants watched a film of a car accident and were then asked how fast the cars were going when they "hit" each other, or "smashed", "collided", "bumped", "contacted". The verb changed the estimated speed by around nine miles per hour. A week later, the participants who had been given "smashed" were more than twice as likely to report having seen broken glass in the film.

There was no broken glass in the film.

This is the misinformation effect, and the important part is the timing: the distortion was introduced after the event, by a single word in a question, and was then reported as a memory of the event itself. Whatever a memory is, it is open to editing at retrieval, and nothing in the subjective experience marks the edited part. The false detail does not feel flagged or hazy. It feels like the rest.

The obvious objection is that these are trivial details in a laboratory film, and that vivid, personally significant events must be stored differently. This is what people mean by a flashbulb memory: the claim that shocking public news is burned in, complete with the room and the light and who said what.

It was tested, because the events keep happening. Researchers collected accounts within a day or two of the Challenger explosion, of 9/11, and of other such mornings, and then went back to the same people years later. The memories were held with enormous confidence, in rich detail, and they were wrong at roughly the same rate as ordinary memories: sometimes spectacularly so, with people confidently describing watching television coverage that did not exist yet. What is special about a flashbulb memory is the confidence, not the accuracy. The two came apart entirely.

The mechanism now proposed is reconsolidation: recalling a memory returns it briefly to an unstable state, and it is then re-stored, incorporating whatever context it was recalled in. If that is right, a memory is not a file that is read. It is a file that is rewritten a little each time it is opened, which means the memory you have rehearsed most is the one that has been through the most edits. Frequently retold stories are not better preserved. They are more worked.

None of this means memory is unreliable in general. It is mostly serviceable, and the gist survives far better than the detail, which is presumably the point, since a system optimised for verbatim storage of everything would be enormously expensive and not obviously more useful. What follows is narrower and harder: that confidence is not a guide, that a detail can be sincerely reported and entirely manufactured, and that anyone whose questions can introduce a detail has more power over the resulting memory than either party is likely to believe.