Remembering can change a memory. But what opens it for editing?

Image: Silver-stained human hippocampus showing myelinated pathways. Alicia Uceda Heras and Alejandro Caraballo Llorente / Wikimedia Commons · CC BY 4.0.

Tell the same family story often enough and eventually somebody objects to the colour of the car. You can picture it. You are certain it was red. A photograph then appears: blue.

Did the act of remembering alter the memory? Possibly. But a changed account is not proof that the stored event was opened, edited and saved again. You may have reconstructed a patchy episode differently, confused the source of a detail, learned the correction as a second version or simply changed what you chose to report.

This matters because one of neuroscience’s most magnetic claims is that every time you retrieve a memory, you rewrite it. There is a real process beneath that slogan—reconsolidation—and some of the animal evidence for it is unusually strong.

The word “every” is where the trouble begins.

Changed recall is not the same as a changed trace

Memory is not a video file waiting to be played. Episodic recall is constructive: the brain rebuilds an event from distributed details, general knowledge, current goals and cues available now. That flexible system lets us extract a gist and use old experience in new situations. It also permits omissions, intrusions and confident errors.

That already means remembering can produce a different result on Tuesday than it did on Monday. Reconsolidation makes a narrower claim. After a consolidated memory is reactivated, the underlying neural representation may become temporarily unstable. To persist, it must be stabilised again; information encountered during that interval may strengthen, weaken or alter it.

The two ideas are related, but they are not interchangeable. A later answer can change even if the original representation was not destabilised. A new safety memory can suppress an old fear. A correction can compete with misinformation. The person can retain both red-car and blue-car versions and become better at selecting one. Behaviour shows which memory won the test, not necessarily what happened to every trace underneath.

So what would count as stronger evidence that retrieval had actually reopened a memory?

A rat, a tone and a very specific interruption

In a landmark 2000 experiment, Karim Nader, Glenn Schafe and Joseph LeDoux trained rats to associate a tone with a mild foot shock. A day later, a single tone reactivated the fear memory. Researchers then infused anisomycin, a protein-synthesis inhibitor, into the lateral and basal amygdala.

The rats still expressed the memory four hours later, but their long-term fear response was impaired at 24 hours. The effect depended on timing and retrieval: the same treatment six hours after reactivation did not produce the impairment, and giving the drug without first presenting the tone left the memory intact. Even a 14-day-old memory became vulnerable when reactivated.

That pattern is difficult to explain as mere distraction during the test. It looks like reactivation placed the fear memory into a time-limited, protein-synthesis-dependent state. Block the restabilisation process and later expression suffers.

It is compelling evidence for reconsolidation in a particular rat fear-learning system. It is not evidence that silently recalling your first day at school automatically rewrites every detail. The drug intervention is invasive, the outcome is a conditioned fear response, and the experiment deliberately engineers the circumstances. Human autobiographical memory is more distributed, layered and difficult to interrogate.

Retrieval may be necessary without being sufficient

If every retrieval destabilised every memory, remembering would be biologically reckless. A useful system must preserve reliable knowledge while changing knowledge that no longer predicts the world.

That creates a stability–plasticity problem: when should the brain keep an old model, and when should it revise it?

One leading answer is prediction error. A memory does more than describe the past; it helps generate expectations. If a familiar cue produces exactly what was expected, there may be little reason to make the memory editable. If the outcome differs, the discrepancy signals that the model might need work.

In a 2013 human fear-conditioning experiment, Dieuwke Sevenster and colleagues manipulated the relationship between original learning and a later reminder. Their results supported prediction error as a condition for propranolol to disrupt the later expression of learned fear. Other animal and human work has also linked incomplete or surprising reminders with memory updating.

But prediction error is not a universal on switch. Its size, content and timing matter; memory strength and age matter; the retrieval cue and what follows it matter. Different studies operationalise “surprise” differently, and researchers often infer destabilisation from the update they were trying to explain. That risks circular reasoning: the memory changed because it destabilised, and we know it destabilised because it changed.

Surprise may create a fork, not just an edit button

The more interesting turn comes when the mismatch is very large.

Suppose your regular café has moved its till by one metre. That is probably the same café with a corrected layout. Suppose you enter the same doorway and find a dentist’s surgery. Treating that as a minor café update would destroy useful knowledge. The brain may instead decide that the evidence belongs to a different situation.

Computational accounts describe this as deciding on a latent cause: did the old state of the world change, or am I now in a new state? A 2025 review proposed a related neurobiological account in which the hippocampus and the brainstem’s locus coeruleus help use prediction-error magnitude to curate memory. Small errors may promote editing of an existing representation; large errors may favour formation of a separate episodic memory.

That is a plausible mechanistic model, not a settled rule with a known numerical threshold. Evidence from human neuroimaging and behavioural studies supports both integration and separation of related experiences, but exactly how the brain draws the boundary is still being worked out.

It also explains why “make the reminder surprising” is incomplete advice. Too little mismatch may leave a memory stable. Enough may open it to updating. Too much may persuade the brain that the new information belongs in another file.

Fear can disappear from a test without being erased

Fear conditioning has supplied some of the most exciting human reconsolidation results—and some of the clearest warnings about reading behaviour too literally.

During ordinary extinction, a person repeatedly encounters a previously threatening cue without the expected bad outcome. Fear declines. Yet the original association may not be gone; a newer safety association can be inhibiting it. Fear can return with time, after an unsignalled aversive event, or when the cue is tested in a different context. Researchers call these spontaneous recovery, reinstatement and renewal.

A widely discussed 2010 study reported that extinction training delivered ten minutes after a reminder prevented the return of conditioned fear, whereas extinction six hours later or without the reminder did not. The result suggested that new safety information had been inserted during a reconsolidation window rather than merely laid on top.

That sounded like the answer. It was not quite.

Other studies produced conflicting results. A 2020 verification report reanalysed the original 2010 data and found that the headline group differences depended on qualitative participant exclusions; with principled alternatives, fear recovery was highly similar between groups. In four direct attempts to replicate a prominent motor-memory reconsolidation result, researchers likewise failed to find the critical impairment effect. In related declarative tasks, retrieval sometimes protected the older sequence instead of making it more vulnerable.

A 2022 review of human and nonhuman studies concluded that reactivation-induced modification has abundant positive evidence but a long history of inconsistent replication. Boundary conditions may be real biology rather than excuses—but a mechanism that works only when delicately tuned is not the same as “recall rewrites memory”.

An emotional response is only one part of a memory

Propranolol studies reveal another hidden complication. The drug blocks beta-adrenergic receptors and can reduce bodily components of arousal. In a 2009 human fear experiment, propranolol before reactivation suppressed the later behavioural expression of conditioned fear and its return.

That does not necessarily mean the participants forgot which cue predicted the shock. A memory can contain factual, contextual, sensory and emotional components that do not all update together. Changing the startle response while explicit expectancy remains available would be clinically interesting, but it is not the science-fiction version of deleting an episode.

The clinical literature is correspondingly unsettled. A 2023 systematic review found 13 randomised trials that added drugs intended to affect extinction or reconsolidation to trauma-focused psychotherapy for post-traumatic stress disorder. Four trials favoured the drug augmentation, seven found no significant advantage and two found smaller symptom reductions with the drug. The agents, procedures and study quality varied too much for a clean overall answer.

Symptom improvement would matter even if reconsolidation were not the mechanism. Conversely, lab evidence that a fear response can be modified does not establish a dependable treatment for complex autobiographical trauma. A useful clinical claim must demonstrate benefit, durability and safety—not merely borrow a plausible molecular story.

Remembering can also strengthen a memory

The viral formulation makes retrieval sound corrosive. Often it does the opposite. Testing yourself is one of the most reliable ways to improve later retention. Reactivation can strengthen a memory, integrate new information, bias it, leave it apparently unchanged or help create a separate episode.

Those outcomes are not contradictions once retrieval is treated as a decision point rather than an overwrite command. The system has to answer at least three questions: Is this old memory relevant? Does the new experience belong with it? If it does, what should change?

This is also why researchers need more than a before-and-after score. Strong evidence for reconsolidation in humans would benefit from an independent marker that a memory became labile, a manipulation delivered inside the predicted time window, no-reactivation controls, and tests designed to reveal whether the old response can recover. Neural or molecular measures should converge with behaviour. For therapy, preregistered trials would need to show that the proposed mechanism mediates a lasting clinical effect.

The stranger truth is a sorting problem

Does remembering change a memory? It can. Recall is reconstructive, and reactivation can sometimes open consolidated memories to modification. But not every retrieval destabilises a trace, not every observed change is reconsolidation, and not every successful update is erasure.

The stronger conclusion is less cinematic and more interesting. Memory has to decide whether present evidence is a correction to the past or the beginning of a new situation. Too rigid, and we keep predicting a world that has changed. Too editable, and every surprise corrupts what we already know.

Reconsolidation may be one of the brain’s solutions to that balance. Prediction error may help open the gate; context and mismatch may determine whether the new information is integrated or kept separate. The failures and boundary conditions are therefore not a footnote to the story. They are the story: a memory system protecting itself from both obsolescence and chaos.

The better question is not “does recall press Save?” It is “how does the brain decide that a new experience belongs to an old memory?”