Funes and the Index
Retrieval is not recollection. A perfect record of the past may be the surest way to have none.
A horse throws a young man in a Uruguayan village and he wakes up crippled and unable to forget. Ireneo Funes, in the story Borges published in 1942, can recall the shape of every cloud he has seen, the grain of every board in a wall, the exact sensations of a particular afternoon — including, if he wishes, the sensations of remembering that afternoon on some later day. The reconstruction of a single day takes him a full day. He invents his own numbering system, giving individual names to numbers up to twenty-four thousand, because a name is more vivid to him than a place in a sequence.
Borges is careful about what this costs, and the cost is the point of the story. Funes cannot think. To think, Borges says in the story’s most quoted judgement, is to forget differences, to generalise, to abstract — and in Funes’s crowded world there are almost nothing but details. The dog seen from the side at three fourteen and the dog seen from the front at three fifteen do not readily fall under one word. Perfect retention has eaten the faculty it was supposed to serve.
The story is usually read as fantasy. It is closer to a case report than that. Alexander Luria, whose account of the case was published in 1968, spent thirty years studying a Moscow newspaperman he called S. — Solomon Shereshevsky — whose memory had no measurable limit, and whose difficulties in Luria’s account run parallel to Funes’s: trouble with abstraction, with metaphor, with recognising a face he had seen in a different light, with reading a passage while a swarm of concrete images crowded in. He had to invent techniques for forgetting, and they did not work well.
The retrieval machine
We have now built the thing that was a curse in the story, and we have built it twice.
The first is the index: a system that stores documents, embeds them, and retrieves the ones nearest to a query. It is the better of the two analogies to Funes, and it works. It also has no relation to the past. An index has no before. Every document in it is equally present, equally available, without the ranking that time performs on a human archive, and without the crucial property that some things have been let go.
The second is the model itself, and here the loose talk begins. A language model’s weights hold compressed statistical structure extracted from an enormous amount of text, with no record of episodes anywhere in them — closer, if an analogy is wanted, to what a lifetime of reading leaves behind in a person than to a diary. The context window is different again: a working buffer holding the current exchange, discarded when the exchange ends. When someone says a model “remembers” what they told it last week, one of two things is true. Either a retrieval system put last week’s transcript back into the window, in which case the machine is reading a document, or nothing of the kind happened and the impression is manufactured by fluency. Neither is recollection, and the word should be used with the metaphor showing.
An index has no before. Everything in it happened at the same time, which is to say at no time.
What recollection turned out to be
The strange thing is that human memory does not work like a record either, and we have known this for close to a century.
Frederic Bartlett’s experiments at Cambridge through the 1920s and 1930s, gathered in Remembering (1932), had English participants recall an unfamiliar folk tale repeatedly over weeks and months. What came back bore little resemblance to a degraded copy. The tale had been rebuilt each time, and rebuilt towards the participant’s expectations: details that made no sense in their culture were dropped or converted into something that did, the story shortened, the causal logic tidied. Bartlett’s word for what did the rebuilding was schema. Remembering, on his account, is construction — performed fresh each time, under the influence of everything the person now believes, with nothing in it that answers to retrieval.
Elizabeth Loftus and John Palmer made the same point sharper in 1974 by breaking it on purpose. Participants watched a film of a car accident, then were asked how fast the cars were going when they hit each other, or smashed into each other. The verb moved the speed estimates. A week later, participants who had been given “smashed” were more likely to report having seen broken glass, which was not in the film. The memory had incorporated the question.
This looks like an indictment and is closer to a specification. A system that rebuilds the past each time under the influence of what it now knows is a system that can revise, integrate, learn from, and finally be finished with an event. The thing that hurt at twenty-five gets re-made every time it is visited, by a person who has changed, until eventually it comes back as information rather than as injury. Funes cannot do that. The record forbids it. He is condemned to the same afternoon at its original resolution for ever, which is one reasonable description of trauma.
The institutional case
The argument so far has been about individuals, where it is easy to be sentimental. It gets harder and more interesting one level up.
Organisations have always had two memories. One is the record — minutes, contracts, tickets, commit logs, the filing system — and it has the properties of Funes: complete within its scope, undegraded, indifferent, unable to tell you which of ten thousand decisions mattered. The other lives in people and is exactly as unreliable as Bartlett would predict: the senior engineer who knows why the module is shaped like that, the account manager who remembers what happened the last time this client was promised a date, the person who says we tried that in 2019 and here is what went wrong. That second memory is reconstructive, self-serving, mortal, and it is the one that actually governs behaviour, because it arrives with a judgement already attached.
What is being deployed now is a very good retrieval system over the first kind, offered as a replacement for the second. It answers what the record says. It cannot answer what the record means, because that sort of meaning accrues to whoever lived through the consequences, and the system was not there. An organisation that routes its institutional memory through an index gets faster and more complete access to everything that was written down, and loses the only mechanism it had for weighting.
The failure mode has nothing to do with the index being wrong. The index has no opinion, and an organisation with no opinion about its own past will repeat it with excellent documentation.
Forgetting is not decay
The word “forgetting” is doing two jobs and they should be separated, because the machines have one and not the other.
Human forgetting is active. Some of it is decay, but the interesting part is not: it is suppression, interference, the reorganisation of an account so that what no longer fits stops being retrievable. It runs on relevance, and relevance is a judgement about what the past is for, made continuously by a system with current purposes. That is why the same event is available at different resolutions to the same person at different times of their life, and why grief attenuates without anyone deciding to attenuate it.
A retrieval system does not forget. It has documents and it has an index, and the only thing resembling loss is eviction — a document removed, a context window overrun, a session ended. None of that is selective in the relevant sense. It has no view about what the past is for, because it has no purposes that outlast a request. What looks like the machine forgetting your conversation is the machine never having had it.
The two systems, then, are not on a scale with each other. One wants to say a model forgets too much and an index too little, as though a well-tuned middle existed. But an index that evicted documents on relevance would still be evicting them for whoever set the policy, and the whole point of the human version is that nobody sets the policy: the forgetting is performed by the same system that has to live with the result.
Proust’s madeleine points at the other half of the difference. What arrives in Du côté de chez Swann is not a retrieved fact but an involuntary flood, triggered by a taste, arriving with a body attached and a mood already fully formed, unbidden and unavailable to command. Whatever that is, no query returns it.
The case against
Start with the charge that all of this romanticises a defect, and that the romance is expensive.
Everything praised above — reconstruction, drift, schema-driven tidying, the merciful softening of what happened — is, in any setting where the past matters practically, a catastrophe. It is why eyewitness testimony has sent innocent people to prison. It is why organisations keep minutes rather than asking people what was decided. It is why the law developed rules of evidence, why science developed the lab notebook, and why every institution that has ever tried to run on recollection has eventually built an index instead. Loftus’s work gets cited, overwhelmingly, by people trying to stop courts relying on human memory at all.
And the sharp distinction between recollection and retrieval may not survive contact with how people actually remember. Andy Clark and David Chalmers argued in 1998 that the notebook in your pocket, consulted reliably and trusted automatically, is functionally part of your memory — that the boundary of the mind is not the skin. If they are right, the shift being described here as a loss is a very old one, running through writing, the archive, the address book and the search engine, and the arrival of a machine that retrieves well is a change of degree in a process that constituted human memory in the first place. On that reading the essay is nostalgia for a purity that the invention of the written word already ended.
What I keep circling is that the two systems fail in opposite directions and we need both, which is not a resolution so much as an admission that the vocabulary is wrong. We say a model remembers and we say a person has a record, and each borrowed word smuggles in the properties of the other thing. Funes’s tragedy lay in having exactly one kind of memory and no capacity to want the other. Quantity was never what ruined him. An institution outsourcing its memory to indexes may well be in the same position, and an institution, unlike Funes, cannot be thrown from a horse and come to.
Sources
- 01Borges, Jorge Luis · 1942 · Funes el memorioso (Funes the Memorious) · La Nación; collected in Ficciones, 1944
- 02Luria, Alexander R. · 1968 · The Mind of a Mnemonist: A Little Book about a Vast Memory · Basic Books; English translation by Lynn Solotaroff
- 03Bartlett, Frederic C. · 1932 · Remembering: A Study in Experimental and Social Psychology · Cambridge University Press
- 04Loftus, Elizabeth F.; Palmer, John C. · 1974 · Reconstruction of Automobile Destruction · Journal of Verbal Learning and Verbal Behavior 13(5), 585–589
- 05Proust, Marcel · 1913 · Du côté de chez Swann (Swann's Way), the first volume of À la recherche du temps perdu · Grasset; the cycle completed posthumously, 1927
- 06Clark, Andy; Chalmers, David J. · 1998 · The Extended Mind · Analysis 58(1), 7–19