More context feels like more knowledge, one of the most expensive illusions in enterprise AI right now.
The instinct is understandable. Context windows keep growing, so the plan writes itself: pour everything in. Every email, every policy, every past decision, every document the company has produced. Give the system total recall and let it sort out what matters. Bigger window, smarter system. That assumption gets more expensive the harder you lean on it.
Past a point, more context does not make a system know more. It makes the system heavier, slower, and worse at judgment. The extra history does not wait quietly in reserve to be recalled. It competes for the model's attention, and it wins competitions it should lose.
The evidence is not soft. Researchers at Stanford and Berkeley documented in 2023 that models attend well to the start and end of a long input and poorly to the middle, with accuracy on a buried fact dropping by more than 30 percent. The industry named the broader effect context rot. As the input grows, attention spreads thin and similar-looking but irrelevant material starts to interfere. Testing across frontier models in 2026 found accuracy falling by 30 to 50 percent as context grew, often well before the stated window limit came into view.
Read that carefully, because it locates the real problem, and the real problem is not storage. Every one of those failures happens at the moment of inference, in what the model is asked to hold in view at once. The harm lives in attention, not in the archive.
Which is why the fashionable prescription is the wrong one. The answer is not to remember less but to keep everything and show the model almost none of it.
You do not delete your history. You store it, durably, and you retrieve the small, relevant, well-ordered slice a given decision actually needs. For a regulated institution this is not even optional. Records retention, legal hold, and discovery all require that the archive stay intact. The compliance officer and the architecture agree for once: the discipline is not forgetting, it is governing what enters the room for each decision.
Borges wrote about a man who could forget nothing. Funes remembered every leaf of every tree he had ever seen, every instant of every day, and was, Borges tells us, almost incapable of thought. To think is to bring a few relevant things into view and leave the rest in the archive. Funes could leave nothing in the archive. Everything was always in the room at once. That is the failure a stuffed context window reproduces. Not too much stored. Too much shown.
Total recall is not a memory. It is a room too crowded to think in.
There is a sharper edge to this, and it cuts the other way. Some things must never leave the room.
When a long-running agent's context fills, it compacts, summarising its own history to make space. In 2026 the security community catalogued what goes wrong, and it is not hypothetical. A compaction step can quietly drop a safety-critical instruction, something as plain as "confirm before acting," and the agent then pursues its goal without the guardrail it started with. The instruction was there. Then it was gone, at the worst possible moment, because nothing marked it as the one thing to keep.
So governed attention has two duties, not one. Decide what to bring into view for a given decision. And decide what is never allowed to fall out of it. The safety rule, the provenance of a number, the exception that changes the answer: those get pinned, protected from the routine summarisation that discards whatever it was not told to keep.
This is design, not default. A rented window compacts however the vendor decided it should. A system you build can be told exactly what to surface and what to protect, with retention, provenance, and expiry set in the architecture itself rather than left to whatever the underlying model does when it runs low on room.
The reframe I would offer any institution is to treat the context window as the scarce resource and the archive as the cheap one. Storage is cheap, and for you it is mandatory. Attention is expensive and quietly decisive, because every item in view competes with every other for the judgment you are paying for. So make each thing that enters a decision earn its place, and hard-protect the few things that must survive every summary.
Keep everything. That is the law and the sound architecture at the same time. Then show the system almost nothing, on purpose, and never let it lose the handful of things it cannot afford to forget.