The Use of Forgetting: Why Systems That Remember Everything Lose the Ability to Judge

1. A Dog and a Grape Leaf

In June 1942, Borges published a short story in La Nación in Buenos Aires; it was later collected in Ficciones. The protagonist, Ireneo Funes, is thrown from a horse and left paralysed in bed — and from then on possesses a memory that never decays.

Borges is restrained about it. Funes can recall the shape of every cloud in the south at dawn on 30 April 1882. He can reconstruct each dream of an entire day. He can identify the binding variations of any Latin volume he has seen. He learned English, French, Portuguese and Latin — not through grammar, but by memorising page after page.

Then Borges adds a judgement: I suspect, nevertheless, that he was not very good at thinking.

And then the sentence that gets quoted most:

To think is to forget a difference, to generalize, to abstract. In the overly replete world of Funes there were nothing but details, almost contiguous details.

(James E. Irby's translation. Andrew Hurley renders the opening as "to ignore (or forget) differences," which is colder still.)

How concrete is the claim? Funes supplies his own example. The dog seen from the side at three fourteen and the dog seen from above at three fifteen are, for him, two distinct individuals. That they share a single name troubles him. He cannot grasp the general term "dog," because in his world no two dogs are the same kind of thing.

He can barely sleep. He tells Borges that his memory is "like a garbage heap."

At twenty-one he dies of pulmonary congestion.

The story is usually read as a fable about memory: remembering too much is a disaster. That reading is not wrong, but it is too quick. It skips the actual question — why should remembering everything make thinking impossible? If memory were storage, more storage should be better. Funes shows that memory is not storage, and that forgetting is not the failure of storage.

2. Forgetting Is a Designed Function

In 2017, Blake Richards and Paul Frankland published a review in Neuron (vol. 94, no. 6) titled "The Persistence and Transience of Memory." Their central claim can be stated counterintuitively: forgetting is not the memory system breaking down; it is an adaptive function of equal standing with remembering.

The mechanisms are not mysterious. Synaptic connections are pruned. Neurogenesis in the hippocampus keeps overwriting old circuits. Every act of retrieval competes with rival memories for the same pathways. The brain is not a hard drive filling up; it is closer to a warehouse that continuously re-indexes — and re-indexing costs you losses.

What the losses buy is three things.

Less interference. If every prior experience were retained at equal weight, judgement about a new situation would be contaminated by a mass of irrelevant precedent. Forgetting is noise reduction.
Generalisation. This is the crux. Only by discarding the differences among instances can a category form. A system that can distinguish every instance has an empty category boundary — it discriminates everything and therefore classifies nothing.
Flexibility. If every past outcome were fully retained, new evidence would struggle to change behaviour. Forgetting is what allows a model to be revised.

Funes lacks all three. He does not know too much; he cannot fold details into a pattern. The dog at three fourteen and the dog at three fifteen remain two events that will not merge, so the word "dog" is, to him, a lie.

The same structure appears earlier in communication theory. Shannon's 1948 paper established that channel capacity is finite, so any transmission must first compress — and compression is, in general, lossy. Information theory offers no universal guarantee of losslessness, only conditional promises for particular sources. Put differently: any transmission system that works must have a built-in decision about what to throw away.

Funes has no such decision. He keeps everything, and therefore transmits nothing.

3. Organisations Turn Memory into an Asset, Then into a Liability

The individual-level conclusion is easy to accept. Its organisational version is harder, because organisations almost never believe they remember too much. Archives, processes, standard operating procedures, compliance records, incident reviews, the comment on every line of a codebase — all of it is accumulated as an asset. And it is one: traceable, transferable, auditable.

The liability does not arise from failing to remember. It arises when what has been remembered acquires normative status.

This is the structure Ray Huang's 1587, a Year of No Significance works through. His central judgement is that the Ming state's problem was governing by moral precept in place of law, which left it unable to manage the empire numerically. The founding institutions of the Hongwu emperor functioned as an unamendable constitution: any adjustment had to argue first that it conformed to the ancestral system, never that it worked better. The Wanli emperor, deadlocked with the civil bureaucracy over the succession, withdrew into decades of passive non-government. Reading him as lazy is the cheapest explanation. The more accurate one is that his room to move had been sealed by memory. The system remembered every precedent, so every change had to prove it was not a betrayal of precedent.

This does not mean the Ming were stupid, nor that the ancestral system was absurd. It means the density of memory determines the space available for adjustment. In a system that treats every precedent as a constraint, the cost of reform rises exponentially with age.

Modern organisations have a plainer version. In a system that has run for twenty years, nearly every line of the codebase corresponds to the patch for some historical incident. What a new engineer sees is not a design but a stratum of accident fossils. At that point remembering has replaced understanding: nobody knows why it is written this way, only that it cannot be changed. The more complete the documentation, the more impossible the change.

Walter M. Miller Jr.'s 1959 novel A Canticle for Leibowitz pushes this to its limit. After a nuclear war, the "Simplification" burns books and kills scholars; a monastery's task is to preserve fragments. The monks copy, venerate and seal every page they find — including a shopping list. The fragments are preserved perfectly; the context is forgotten entirely. The consequence is that the preservers need centuries before they can read what they have been preserving.

That is the worst kind of forgetting — all the details present, the pattern gone.

4. The Counterargument: Forgetting Is Never Neutral

At this point a dangerous conclusion is easy to reach: the more you forget, the healthier you are, and organisations should periodically wipe their memory. In politics that conclusion is catastrophic.

Tony Judt returns repeatedly in Postwar to a single phenomenon: in rebuilding, European states generally chose not to ask what had happened during the war. The German case is recorded in finer grain in Aftermath — people discussed the ruins, the hunger, their own expulsion from the east, but not what they themselves had done. Functionally, that silence helped: it allowed societies to start rebuilding without first completing a reckoning on the rubble.

But the bill was not cancelled, only deferred — and it accrued interest. From the 1960s onward a generation began asking about its parents' roles, and the conflict reopened. By then there were fewer witnesses, evidence was harder to verify, and argument slid more easily into position-taking. A postponed reckoning does not become gentler; it becomes harder.

The deeper question is who holds the right to forget.

A person forgetting his own pain is exercising a capacity. An institution forgetting its own conduct is claiming an exemption. Mechanically the two are identical — both stop certain records from participating in present judgement — but their moral standing is opposite. Conflating them is the easiest mistake available to the argument that forgetting is good for us.

The technology of record has always been a technology of power. One of Harari's central claims in Nexus is that information networks build order through records and maintain order by rewriting them; what binds a network is often not truth but a fiction people are willing to cooperate around. Deleting a page of archive and killing a person can be equivalent in their effect on order, while only the first avoids the charge of murder.

So "forgetting is beneficial" has to be qualified. It holds only when those who are forgotten can take part in authorising the forgetting.

5. Three Kinds of Forgetting, Three Tests

Pulling the threads together yields three kinds of forgetting with entirely different properties.

First, abstractive forgetting: details dropped, pattern kept. This is the capacity Funes lacks and the expert possesses. Someone who has learned a craft cannot explain the steps and still gets it right; someone who has learned a language cannot recite the rules and still knows when a sentence is wrong. This kind is beneficial: it compresses experience into portable judgement.

Second, archival forgetting: details kept, pattern lost. The monastery's fragments, the codebase of accident patches, a discipline with data but no theory. Nothing is discarded, but the "why" is. Its signature symptom is knowing everything and being unable to judge — every decision requires reopening the archive, and the archive has no priorities.

Third, erasive forgetting: records destroyed to escape accountability. This one is not controversial. Its defining feature is irreversibility: not "I cannot recall for now" but "it can never again be verified."

Three tests follow, for deciding whether a given act of forgetting is a capacity or a failure:

Grain. Are instances being dropped, or criteria? Dropping instances is abstraction; dropping criteria is disablement. The first makes you faster, the second makes you blind.
Authorisation. Who decides? Are the forgotten present? A unilateral decision to forget, however well argued, should be treated as the third kind until proven otherwise.
Recoverability. Can what was dropped be recovered? Forgetting with cold backups, independent audit and off-site copies is a different act from destruction. Only what is technically recoverable can be "judged later."

Then the limit. None of this argues that one should forget whenever possible. Medical records, judicial files and raw scientific data must be kept indefinitely, precisely because their value lies in re-examinability — a finding that cannot be re-checked is not a finding. And the cost of forgetting is always paid by the future, which has no vote. That is exactly why forgetting looks so cheap inside organisations: it saves storage and explanation today, and spends judgement ten years out.

Funes' tragedy is not that he remembered too much. It is that his memory was not his to command — he could not choose what to drop, and therefore could not choose what to keep. The real line does not run between remembering and forgetting. It runs through who decides.