A We tend to speak of memory in the language of storage. Recollections are “filed”, “retrieved”, occasionally “lost”; the brain is imagined as an archive of some kind, imperfect only in that its cataloguing system deteriorates with age. The metaphor is intuitive and almost entirely wrong. Fifty years of experimental work suggest that remembering is not the recovery of a stored object but the construction of a new one, assembled each time from fragments, expectations and general knowledge — and that the act of remembering alters the very thing remembered.

B The first systematic challenge to the archive model came from Frederic Bartlett at Cambridge in the 1930s. Bartlett read English participants a Native American folk tale, The War of the Ghosts, whose structure and supernatural logic were unfamiliar to them, then asked them to reproduce it repeatedly over days, weeks and in some cases years. The reproductions did not simply fade. They changed in a consistent direction: unfamiliar details disappeared, supernatural events were rationalised into ordinary ones, and the narrative gradually acquired the shape of a conventional English story. Bartlett proposed that participants were not recalling the tale but reconstructing it around a schema — an existing mental framework of how stories generally work. Where memory was silent, the schema supplied a plausible substitute, and participants could not distinguish the supplied material from the original.

C For four decades this work was admired more than it was applied. That changed with a series of experiments by Elizabeth Loftus in the 1970s. In the best known, participants watched a film of a car accident and were then asked how fast the vehicles were travelling when they “contacted”, “hit” or “smashed into” each other. Estimates of speed rose with the violence of the verb. A week later, those who had been given “smashed” were more than twice as likely as the others to report having seen broken glass in the film. There was no broken glass. A single word, embedded in a question, had inserted an object into a memory of an event. Loftus and others went on to show that entire episodes — being lost in a shopping centre as a child, spilling punch at a wedding — could be implanted in a substantial minority of adults through nothing more than suggestion and repeated questioning.

D The mechanism became clearer with the discovery of what is now called reconsolidation. Neuroscientists had long held that a memory, once fixed in the brain over hours or days, was essentially stable. Work in the early 2000s, initially on rats and later on humans, indicated instead that recalling a memory returns it temporarily to an unstable state, in which it must be restabilised — reconsolidated — before it settles again. During that window the memory is open to modification. Anything present at the moment of recall, including the emotional state of the person recalling and the phrasing of the question that prompted them, may be incorporated. The counter-intuitive implication is that the memories we revisit most often are not our most reliable but our most edited.

E These findings have had their sharpest consequences in the courts. Eyewitness identification was for most of the twentieth century regarded as the strongest evidence available. Analysis of the first several hundred convictions overturned by DNA evidence in the United States found that mistaken eyewitness identification had contributed to roughly seven in ten of them. Confidence was no protection: witnesses who were certain were, in a number of studies, no more accurate than those who were hesitant, though juries treated them as far more credible. Procedural reforms followed — administrators who do not themselves know which member of a line-up is the suspect, witnesses shown faces one at a time rather than together, and a confidence statement recorded at the moment of first identification rather than at trial, by which point months of questioning have usually inflated it.

F It would be a mistake, however, to conclude from this that memory is simply defective. Reconstruction is not a design flaw but the price of a system optimised for something other than accurate playback. A memory that recorded every detail of every experience would be enormous, slow to search, and largely useless: what an organism needs is not a record of the past but a guide to the future. By storing the gist of experience and reassembling the particulars on demand, the brain produces something flexible enough to be recombined into predictions and plans. There is direct evidence for this dual function. Patients with damage to the hippocampus, who cannot form new episodic memories, also prove unable to imagine novel future scenes in any detail; asked to describe a hypothetical beach, they produce fragments without spatial coherence. Imagination, it appears, runs on the same machinery as recollection.

G This reframing changes what we should ask of memory. The relevant question is not whether a recollection is accurate in the way a recording is accurate, but whether it is accurate enough for the purpose to which it is being put. For navigating a life, the reconstructive system performs remarkably well. For determining whether a particular person stood on a particular street on a particular evening eighteen months ago, it performs badly, and no amount of sincerity on the part of the witness can improve it. The error has never been in the memory system. It is in our confidence about what that system was built to do.