Explainer

Why long AI chats get worse: context rot, explained

You've felt it: an AI chat that was sharp an hour ago is now forgetting your rules, repeating itself and drifting. It's not your imagination, and it's usually not a secret model downgrade. It's a real, mechanical effect — and once you understand it, you know exactly what to do.

What context rot is

A model doesn't "remember" your conversation the way a person does. On every single reply, it re-reads the whole thread from the top — all your messages, all its answers, everything you pasted — and produces the next response from that entire pile. The pile is called the context window, and it has a fixed size.

Context rot is the observed pattern that as that window fills up, answer quality degrades — and not gently. Early on, the model handles your task cleanly. As the thread grows long, its attention is spread across a wall of text, and the signal you care about (your latest decision, that one constraint) gets buried under everything else.

Why it happens

Two things stack up:

Dilution. The more text the model has to attend to, the less weight any single instruction tends to get. That "keep it short" you set early on is now competing with a long wall of other text.
Recency bias. Models lean on what's most recent. Important things you established early — the ones you assume are "locked" — are exactly the ones most likely to fade.

Put together: the longer a chat runs, the more likely it is to forget precisely the decisions you made at the start and never repeated.

How to spot it in your own chat

You can usually feel rot before you can name it. The tells:

It forgets rules you set. You told it something many messages ago; now it acts like you never did.
You keep repeating yourself. You've restated the same constraint two or three times.
It repeats itself. Answers come back nearly identical to ones you already got.
It contradicts settled decisions. Something you locked in earlier quietly comes undone.
It drifts. Replies get vaguer and less anchored to what you're doing.

The quick test: model or thread?

Before you blame the model, run this. Open a brand-new chat and ask the same question you're stuck on. If the fresh chat nails it and this conversation keeps fumbling, the thread is rotten — not the model. If the fresh chat struggles too, the issue is genuinely the task (or a real model change), and restarting won't help. This single check stops you wasting an afternoon fighting the wrong problem.

You can't un-rot a conversation

There's no button that makes a saturated chat forget its own bloat. Trimming messages doesn't help — the model still re-reads what's left, and the important early context is already lost in the noise. The only reliable fix is to move your work into a fresh chat that starts clean.

What to do about it

Don't nurse a rotting thread — hand it off. Give a fresh chat a short briefing: the goal, the decisions you locked, the constraints it kept dropping (list these explicitly — they're the first casualties of rot), and where you left off. A clean chat with a tight briefing is faster and sharper than a bloated one, and it won't relitigate what you already decided.

Building that briefing by hand works but is tedious, and the constraints you were most tired of repeating are the ones you'll forget to include. That's the gap Uncook fills.

The fix that works

Move your work to a fresh chat with Uncook

Free & unlimited. No account.

Paste a share link to your ChatGPT or Claude conversation. Uncook reads it end to end and writes a clean reboot prompt — goal, locked decisions and the constraints the model kept dropping, all carried across. Skim it, paste it into a fresh chat, and you're back to sharp.

Uncook my chat →

Honest about your data: pasted text is analyzed in your browser; a share link is fetched once through our server to read the conversation, then discarded — never stored, never used for training. A share link makes the chat viewable by anyone with the URL; un-share it once you're done.