Human memory shifts the goalposts. It selects, loses and reconstructs. An episode that seemed insignificant to us for years can become amusing; a remark we overlooked can acquire importance much later. Above all, the past is not always available. We may search for it without finding it, until a smell, a voice or a taste opens something up again. For Proust, a madeleinette dipped in tea is enough: from the sensation, the world of Combray resurfaces, with Aunt Léonie. The memory does not arrive like a document requested from an archive. It comes through an encounter between what is happening now and what remains of then.

This instability makes memory fallible. Elizabeth Loftus has shown how even the way a question is phrased can influence an account of an event. Remembering is no guarantee of recovering what happened, and feeling certain offers no protection against error. The present takes part in reconstructing the past: later information can enter memories and alter them. Memory preserves traces, but the story we draw from them does not necessarily stay the same.

Yet this imperfect memory has useful properties. Neuroscientists Blake Richards and Paul Frankland have described how forgetting can foster better decisions in changing environments: it reduces the weight of outdated information and allows us to generalise beyond individual episodes. If every experience remained bound to all its details, it would be harder to recognise what different situations have in common. Losing something can help us apply what we have learnt elsewhere.

In evolutionary terms, accuracy of recall is a capacity to be balanced against other needs: adapting, recognising patterns, responding to change. This does not make all forgetting advantageous. It means that an effective memory is not necessarily one that retains everything. Daniel Schacter and Donna Rose Addis have also shown that remembering the past and imagining the future involve partly shared processes and circuits. Elements of experiences can be recombined into scenarios never lived through. The same flexibility that leaves memory open to distortion helps us imagine new possibilities. It is not forgetting alone that produces creativity: it is a system capable of preserving, selecting and reorganising.

An artificial memory can be effective precisely because it performs a different task. We ask a digital system to keep data unchanged. Its stability allows us to compare versions, verify claims and retrieve details we have lost. It is useful that a document does not follow the fluctuations of our interpretation. But it preserves a trace, not the entirety of what happened: a recording can reproduce the words of a conversation without exhausting what that conversation meant to the people involved.

Nor is artificial intelligence the same thing as an archive. A language model generates responses through what it has learnt and, when it has access to search tools, through retrieved documents. Patrick Lewis and colleagues formalised this combination by distinguishing the knowledge embedded in the model from the external memory it can consult. The document can remain unchanged while the account produced changes. Artificial networks can forget too: James Kirkpatrick and colleagues have studied the loss of previous skills during new learning. The distinction, then, is not between a human being who forgets and a machine that always preserves everything.

The difference concerns what each form of memory makes possible. Digital stability is an advantage when verification is needed; on its own, however, it does not offer the flexibility of human memory. A perfectly preserved store does not know which details have ceased to matter, which experiences might be brought together or which meanings need to change. These capacities require further processes of selection and reworking. A machine, too, can incorporate them, but it does not possess them simply by storing information. Precision in preservation does not automatically encompass the ability to face a future different from the past.

For us, moreover, the past is lived. Endel Tulving described episodic memory in terms of the ability to recognise oneself in the remembered experience: knowing that something happened and remembering being there are different conditions. A machine can hold extraordinarily detailed information about our lives; this alone does not demonstrate that it has the experience of having lived them. The quantity of data does not settle the question of the subject to whom a memory belongs.

Nietzsche attributed an active function to forgetting: making room for the present, preventing every experience from continuing to occupy consciousness. This is a philosophical perspective, but it brings a concrete problem into view. The past can be preserved and yet lose its hold on us. We can remember a failure without continuing to define ourselves solely by that failure.

Human memory leaves us open to error, but its capacity to rework things contributes to our ability to change. The stable memory of machines helps us verify what we have forgotten or rewritten. Its value lies partly in this resistance. The difficulty arises when we expect better preservation to mean better understanding, or better living. A complex future requires us to use the past without having to repeat it. And a memory that loses something may, for that very reason, leave us room to become something else.