The analogy

Imagine a desk on which you keep piling up sheets without ever removing any. At first you find everything in a flash. After an hour of piles, to answer a question you fish out the wrong sheet, or you keep fishing out the one on top because it's the easiest to grab. The important documents are still there, but they've ended up buried under the others.

The AI in a very long chat has that desk. The whole discussion passes before it with every answer, and the more the pile grows, the more the pieces that matter risk staying at the bottom. From there the two typical consequences: it repeats what's within reach and loses sight of the starting point.

How it really works

The AI doesn't have infinite attention. At each turn it rereads all the text of the conversation, and when the text is enormous its ability to give the right weight to every part gets diluted. The important instructions, if you gave them at the start, end up drowned in the middle of so much else: it's a phenomenon that those who work with these tools call context rot, the "rotting of the context," that is, the degradation of the answers as the window fills up. The AI then tends to anchor itself to what's most recent or most repeated, and it repeats. Or a secondary topic, mentioned in passing, takes over and sends it off the rails. Making it all worse are the vague requests, which leave room for wandering.

What you can do in practice

  • Open a new chat for each new topic. Don't drag a subject into a conversation that's already long about something else.
  • For a long job, every so often ask for a summary of the key points, open it in a fresh chat and paste it in: you restart with a clear desk and only the sheets you need.
  • Repeat the current goal toward the end of the message, where the AI gives it more weight, instead of counting on what you wrote twenty exchanges earlier.
  • Make one request at a time and be specific: less room for the vague, fewer digressions.
  • If it goes into a loop, tell it flat out: "don't repeat yourself, just give me the point that's missing."

A common misconception

People believe a very long chat is better, because "the AI has more context and therefore understands more." Beyond a certain threshold it's the opposite. The more you fill the window, the more the important details risk staying buried or losing weight, and quality drops instead of rising. The infinite conversation doesn't make the AI more informed: it makes it more confused. Better a focused chat and, when needed, a clean restart with the summary.

Frequently asked questions

Is one mega-chat or many short ones better?

Many, focused. One conversation per topic keeps the desk tidy and the answers on point. Piling everything into a single infinite chat is exactly what triggers repetitions and digressions.

How do I restart without losing the work done?

Ask the AI to summarize the key points and the decisions made, copy that summary and paste it into a new chat as a starting point. You take the gist with you and leave behind the noise that was clogging the conversation.

Is it the fault of a poor model?

No, it happens to the best ones too: it depends on the context window filling up, not on the quality of the model. The more capable models hold up over longer conversations before giving in, but the mechanism is the same for all.

If it repeats, does it mean it didn't understand the question?

Not necessarily. Often it understood perfectly well, but it's overloaded and clings to what it has already said. The proof is simple: bring the same exact question into a new, clean chat. Nine times out of ten you get a focused answer. It wasn't misunderstanding, it was clutter.