In 1942 Jorge Luis Borges wrote a short story about a man named Funes who, after a fall from a horse, could forget nothing. Every leaf on every tree he had ever seen, every shade of every cloud, every sentence spoken to him, stayed with him in total, undifferentiated detail.
Borges did not write this as a gift. Funes could not generalize. A dog seen from the side at three fourteen was, to him, a different object from the same dog seen from the front at three fifteen, because his mind kept every instant separate and none of them collapsed into a category called “dog.” He could not think, because thinking requires forgetting most of what you know in order to use the part that matters. Total memory did not make him wise. It made him unable to reason at all.
That is not just a literary idea. In 2017, the neuroscientists Blake Richards and Paul Frankland published a paper in Neuron arguing that the biological purpose of memory is not to preserve the past. It is to optimize decisions in a changing, noisy world.
Forgetting, in their account, is not a failure of the system. It is a feature the system runs on purpose. An animal that kept every detail of every past event, weighted equally, would drown in irrelevant specifics and lose the ability to generalize from experience to the next decision. Persistence and transience work together: transience discounts information that is no longer useful and prevents the brain from overfitting to events that will not repeat. A mind that remembers everything is not a stronger mind. It is a mind that has stopped doing the actual work of memory, which is judgment about what still matters.
The same failure shows up in machines now, and it is not theoretical. Researchers studying large language models with very long context windows found a consistent pattern: when relevant information sits in the middle of a large amount of stored text, accuracy drops by more than thirty percent compared to when the same information sits at the start or the end. The model has the information. It technically remembers. It just cannot find the signal inside the noise it is carrying.
The same problem shows up one level up, in retrieval systems built to search across large stores of saved information: as the size of the index grows, accuracy on complex questions falls, not because the system forgot, but because it now has to search through everything it never should have kept in the first place.
And when a consumer product tried to solve this by having an assistant remember everything about a user automatically, the result was not a more helpful assistant. Users described it as building a detailed but wrong profile of them over time, a kind of memory rot where stale assumptions and outdated preferences quietly poisoned later answers. Enough people found this bad enough to turn the feature off.
I build systems for a living, and the one I use every day is deliberately built the other way. It does not store everything I say as equally true and equally permanent. It keeps memory in separate, walled scopes, so a fact learned while I am working on one thing cannot leak into a different context where it does not belong; an unlisted scope is not deprioritized, it is structurally invisible to the system, the same way a locked room is not just unlikely to be entered, it cannot be.
Nothing becomes a permanent fact the first time I say it. A new piece of information sits as a candidate and has to survive being checked against reality more than once before it is trusted as durable; most things I say in passing never clear that bar, and they simply stay candidates, neither remembered as truth nor deleted, just not yet believed.
When something does become a standing fact, it gets exactly one home, not five slightly different versions scattered across different files that will eventually contradict each other, because a mind that stores the same fact in six places is a mind that will eventually disagree with itself. And when the system decides what to bring back into a conversation, it does not dump everything it has ever stored at me. It only surfaces something if it clears a real relevance threshold, and only displays it if that relevance is strong enough to be worth interrupting me for. Most of what the system technically knows about me never comes up, because most of it is not relevant right now, and a system that cannot tell the difference between relevant and merely known is not remembering, it is hoarding.
The most important boundary is not even about relevance. It is about what must never be allowed to connect at all. Some of the work I do involves information that belongs to someone else and can never be allowed near anything else I know, under any circumstance, no matter how useful the connection might seem in the moment.
The right way to guarantee that is not a rule the system promises to follow. It is a wall the system cannot get around, because the confidential information lives in a separate, physically isolated environment from everything else, so there is no shared space left in which a mistake could even happen. A promise can be broken by a bad prompt, a bug, or a moment of bad judgment. A structural wall cannot. If you actually care about a boundary, you build it so that crossing it is not merely against the rules, it is not possible.
None of this makes the system perfect, and it should not pretend to be. I know of at least one real bug in it: a fact that had been correctly promoted to permanent memory later got quietly overridden by a narrower, more accurate decision, and the older fact kept surfacing anyway for a while, because nothing was checking whether the standing memory still matched the newer judgment.
That is not proof the design was wrong. It is proof that even a system built specifically to avoid remembering the wrong things has to keep working at it, the same way a person who has decided to be a careful thinker still has to actively catch themselves holding onto a belief that new evidence has already replaced. Curatorial memory is not a setting you turn on once. It is an ongoing discipline, for a system and for a person, and it fails quietly if nobody keeps checking it.
This is also, I think, why the more serious AI labs are moving toward memory you can see and edit rather than memory that happens to you automatically and invisibly. That is not a small design choice. It is an admission that total, automatic recall was never actually the goal.
The goal was trust, and trust does not come from a system that remembers everything about you. It comes from a system, or a person, that remembers the right thing at the right moment and has the judgment to leave the rest alone. Think about the people you actually trust with your history. It is never the person who can recite every word you have ever said to them. It is the person who remembers the one thing that matters, brings it up at the right time, and has the sense to forget the rest. That is not a lesser form of memory. It is the only form of memory that was ever actually useful.
I write about AI, systems, markets, and the incentives hiding underneath them. Get the next essay by email.
Unsubscribe any time. Privacy.