What is a memory made of?

The patient known as H.M., after bilateral removal of the medial temporal lobes in 1953, could no longer form new conscious memories of events. The operation left his existing technical skills intact, and in 1962 Brenda Milner found he could still learn a new one: he learned to trace a star seen only in a mirror and kept the skill across three days, yet at the end had no recollection of having done the task before.

That dissociation opened the modern study of memory as several systems rather than one: conscious memory for facts and events depends on the medial temporal lobe, while skills, habits and simple conditioning depend on other structures.

Neurons active during an experience can be genetically tagged and later artificially reactivated. Doing so makes a mouse behave as though it is recalling that experience, in a place where the experience never happened. The same techniques can suppress a specific memory or attach a false one. Nearly all of it is rodent work.

This is as close as neuroscience has come to holding a memory in its hand.

Retrieving a consolidated fear memory returns it, in rats, to a labile state in which it must be rebuilt to persist. Blocking protein synthesis in the amygdala during that window produces amnesia on later tests, whether the memory was a day old or a fortnight. In rats, at least, remembering something can change it. Whether the same happens in people is argued: a 2018 review of the human studies finds much that is consistent with it, and too many inconsistent results and rival explanations to conclude it.

Changing a single word in a question about a filmed crash (“smashed” versus “hit”) raises the speed witnesses report, and a week later it more than doubles the number who say they saw broken glass that was never in the film. Suggest a childhood event that never happened, with a relative’s account behind it, and in a 2023 replication about a third of people were coded as remembering it. How many of those amount to a full memory is argued: a systematic review of such studies puts it at about 15 per cent at most.

The dominant account places memory in patterns of synaptic strength across populations of neurons. Engram studies find changes in the intrinsic excitability of cells working alongside it. One hypothesis stores very long-term memories in the pattern of holes in the matrix that wraps neurons, and work in mice gives astrocytes a part in forming remote memories. The synaptic account is well supported and probably incomplete.

Sleep benefits memory, and in rats hippocampal activity patterns from waking are replayed during sleep. How large the benefit is has been revised down: a 2021 review, one of whose authors wrote the 2013 review this passage cites, reports key findings that could not be replicated and an effect smaller and less robust than was assumed. Sleep appears to favour some memories over others, such as those that were emotional, rewarded or flagged as needed later, but how the sleeping brain does the choosing remains largely uncharacterised.

A person can retrieve a specific afternoon from decades ago without any obvious search. The mechanisms in the sections above are about how a memory is held; how a lifetime of them is indexed so that one comes back rather than another is a different question, and this page has found no account of it to point at.

Forgetting is not simply decay. Work in flies and mice has found mechanisms that actively erase memory traces, and one proposal holds that forgetting helps learning, by clearing outdated detail so that what is kept generalises. Where the line falls between a memory that is lost and one that is merely inaccessible is not known, and may not be a well-formed question.

Patients with hippocampal damage, asked to imagine a commonplace scene they have never been in, produce fragments with no setting to hold them. Four of the five tested that way were impaired; the fifth was not. Its authors took this to mean that the same structure builds the past and the possible. Not every study agrees: a 2010 study of five patients with hippocampal damage found their ability to imagine the future intact. Whether memory is a system for recording what happened, or a system for simulating what could happen that uses the past as raw material, is genuinely open.

If the point of storing the past is to anticipate the future, then reconstruction is not a flaw in the system. It is the system working. Distortion, blending and updating would be expected consequences of a mechanism optimised for usefulness rather than fidelity.

A single memory may exist as overlapping patterns across multiple regions, which would explain why memories degrade gracefully rather than disappearing, and why decades of lesion experiments in rats never found a place where a specific memory was kept.