AI memory needs to forget

Jorge Luis Borges published a story in 1942 about a young man in Uruguay who falls from a horse and wakes up unable to forget anything.

Ireneo Funes remembers every leaf on every tree he has ever seen, every perception of every leaf, every memory of every perception. He spends his remaining years in a dark room, because light produces more detail and detail produces more memory. He invents a numbering system in which every integer has its own arbitrary name, because grouping integers under a single numeral offends him with its imprecision.

Borges' narrator concludes that Funes was not very capable of thought. To think is to forget differences, to generalise, to abstract. Funes cannot generalise, because generalisation requires discarding, and he discards nothing.

The technology industry is building Funes. It is calling this a feature.

In 1991, John Robert Anderson and Lael Schooler published a finding that surprised the field. Human forgetting rates are not a failure of biological storage. They are an optimal fit to the statistical structure of information need in the real world.

Take any class of information and measure how often it is needed again over time. The probability declines, and the decline follows a power law. Plot this against the human forgetting curve. They match. Memory is tuned to the world, not to fidelity.

Robert Bjork added the mechanism: retrieval strength and storage strength are separable. Losing retrieval access is not the same as losing storage. Forgetting is how the system maintains utility, clearing the path for what is actually needed next. The Ebbinghaus forgetting curve is not a leak. It is a calibration.

A system that never forgets is a miscalibrated one. It will surface, with equal confidence, a preference you mentioned once and then reversed, a phase of your life you have spent years leaving, a version of a relationship that has moved past itself.

The deeper issue is identity, not efficiency.

People change partly by being allowed to be forgotten. A fresh start works not because of the geography but because nobody in the new place has a model of who you were. You are not continuously being read back into a version of yourself you are trying to leave.

A system holding every version of you at equal fidelity will predict you from the average of everything you have ever been. It has no mechanism to weight the present more heavily than the past. It will keep pulling you toward who you were, not because it intends to, but because that is how a system with no decay function behaves.

This is a design problem, not a storage problem. Forgetting must be engineered: weights that decline according to need-probability and distance from who the person is becoming. Not a deletion policy triggered by a GDPR request. An active decay function, engineered to behave the way biological memory actually behaves.

Helen Nissenbaum's contextual integrity framework holds that privacy is not secrecy. It is about appropriate information flow within a context. Something disclosed at 2am in distress has a context. That context is not compatible with the same information surfacing at 11am in a productivity framing. Contextual integrity has been violated even if no data has left the building and no policy has been broken.

Almost no AI memory system models context at all. They model facts. Flattening context is what destroys integrity, and the architecture does it by default.

Borges did not say whether Funes suffered. He implied it throughout. The question for anyone building AI memory is not whether to retain. It is what kind of system they are building: one that serves the person's present self, or one that pins them to their own history.