The archive and the algorithm

An archive is one of the strangest things a society builds, because nothing in the present demands it. The records it keeps are, almost by definition, the ones no one currently needs. The register of a transaction long settled, the minutes of a meeting whose participants are dead, the file on a decision that was made and never questioned — these cost money to store and effort to maintain and return nothing to the institution that keeps them, until the rare day, often decades out, when someone needs to know exactly what happened, and the archive is the only thing in the world that can say. An archive is a bet that the past will be asked about. It is memory kept on purpose, against an occasion no one can name.

What is remarkable is that human institutions made this bet at all. Keeping records is not the natural state of an organization; forgetting is. The default is to clear the desk, close the matter, move the resources to the next problem. An archive is a deliberate refusal of that default — a discipline imposed against the institution's own convenience, sustained by people who would never benefit from it, on behalf of questioners who did not yet exist. It is, in the most precise sense, a moral technology: a structure that encodes an obligation the present owes the future, and keeps that obligation even when no one is watching and nothing is at stake.

The instinct of the algorithm

An algorithmic system has the opposite instinct, and it has it for reasons that are not careless but constitutive. The whole logic of optimization is to improve, and to improve is to replace — to overwrite the old state with a better one, to discard the inputs once the output is computed, to treat yesterday's configuration as superseded waste. A well-engineered system holds only what it needs to do the next thing. Its memory is working memory, kept exactly as long as it is useful and not a moment longer. There is no malice in this. It is what efficiency looks like. But it is the precise inverse of the archival instinct, and the inversion has a consequence the engineering does not register: a system built this way is structurally incapable of being asked about its past, because it has kept nothing the past would need.

This is the collision at the center of automating consequential decisions. We are handing the work of judgment to systems whose deepest design instinct is to forget, and asking them to perform a function — deciding things that affect people — that has always depended on the opposite instinct, on the willingness to keep. The institutions we are replacing kept archives. The systems replacing them keep only what optimization requires, which is to say, almost nothing about the individual decision once it is past. We have not just changed who decides. We have changed whether the deciding can be remembered.

The archive was never about the past. It was the present admitting that the future would have questions, and choosing, in advance, to be able to answer.

An archival discipline, restored

The answer is not to make algorithms nostalgic, and not to abandon optimization, which is doing real and valuable work. It is to recognize that an archival discipline is a separate function from the work itself, and that it has to be added deliberately, because nothing in the work will produce it as a byproduct. The old institutions did not get their archives for free either. They built them as a discipline laid alongside the daily work, staffed and maintained on purpose. The task now is the same task: to lay an archival discipline alongside automated decision-making, capturing what a future questioner will need at the moment the decision is made, before the optimizing instinct overwrites it.

And the discipline is selective, as a good archive always was. An archive is not a hoard; it does not keep everything, indiscriminately, forever. It keeps what a future inquiry will plausibly need — in the case of a decision, the evidence that was actually consulted, the rule that was actually in force, the state from which the conclusion actually followed. A Decision Receipt is an archival instrument in exactly this older sense: not a log of everything the system did, but a deliberate preservation of the things that make a particular decision answerable later. It is the archivist's judgment applied to the algorithm's output — the choice of what to keep against the day someone asks.

To call this nostalgia is to misread what an archive ever was. The archive was not a sentimental attachment to old paper. It was the operational form of accountability across time — the recognition that an institution which cannot produce its past cannot answer for it, and an institution that cannot answer for its decisions is not really accountable at all, however well it behaves in the present. Restoring an archival discipline to automated decisions is not a backward-looking gesture. It is the forward-looking one. It is how a system built to forget becomes a system that can be held to account, which is to say, the only kind worth trusting with decisions that matter.

The algorithm optimizes the present. The archive answers to the future. A society that wants its automated decisions to remain answerable cannot choose between the two. It has to build the second back in, on purpose, alongside the first — because the future will have its questions, and the only institutions that survive them are the ones that chose, in advance, to be able to answer.

— Dispatches · Summit Cognitive