Why do long chats get worse, and when should you start a new one?

Almost everyone who uses ChatGPT, Claude or Gemini for real work has seen it happen. The first twenty messages go well. Then the assistant starts ignoring a rule you set at the beginning, mixes up two versions of a draft, or brings back an idea you rejected an hour ago. Nothing is broken. You have reached the practical limits of how these models read a conversation.

The fix is simple once you understand the cause: know the warning signs, and move to a fresh chat with a short, clean summary of what matters.

What the model actually sees

A chatbot does not remember your conversation the way a person does. Each time you send a message, the app sends the model the conversation so far, or as much of it as fits, and the model writes a reply based on that text alone.

That text is measured in tokens, the pieces of words a model reads. OpenAI's rule of thumb is that a token is about four characters of English, or roughly three quarters of a word, so 100 tokens is about 75 words. Every model has a context window, the maximum number of tokens it can take in at once. When a conversation grows past it, the app has to drop or compress older parts, and what was said at the start may simply no longer be there.

Why quality drops before you hit the limit

Even when everything still fits, more text is not always better. In a 2023 study titled "Lost in the Middle," a team led by Nelson F. Liu gave language models long inputs with one useful piece of information hidden at different positions. The models usually did best when that information sat near the beginning or the end, and worst when it was in the middle. That held even for models designed for long inputs.

A long chat is exactly that kind of input. Your original instructions are now somewhere in the middle, surrounded by drafts, corrections and side questions. The model also sees every abandoned version of your work, and it can't always tell which one you settled on.

Chat products have noticed. In February 2023, Microsoft capped the number of turns in a Bing chat session after finding that very long sessions could confuse the underlying model. Starting over is a reset the system itself sometimes needs.

Signs it is time for a new chat

One sign alone may just need a firm reminder. Two or more usually mean a fresh start will be faster.

How do you carry the important parts into a fresh chat?

Ask the old chat to write its own handover, then check it before you use it. A summary written by the model is quick, but it can leave things out or keep decisions you reversed, so you remain the editor. Here is a worked example for someone planning a community newsletter:

We are going to continue this work in a new conversation.
Write a handover summary I can paste there. Use these headings:
1. Goal: what we are making and for whom.
2. Decisions: everything we agreed, as short bullet points.
3. Rules: tone, length, format and words to avoid.
4. Rejected: ideas we discussed and dropped, so they are not suggested again.
5. Current version: the latest text in full.
6. Next step: the one thing to do first.
Keep it under 400 words, except the current version.

Read the result line by line. Delete anything that is no longer true, add anything missing, and make sure the "Current version" really is the latest. Then open a new chat and start it with one line of framing, such as "Here is a summary of earlier work. Read it and confirm the goal and rules before we continue," followed by the summary.

Two habits make this easier. Keep a running notes file outside the chatbot where you paste key decisions as you make them, so you never depend on the model's memory. And put your most important rules at the top of the new chat, where the research suggests they are most likely to be followed. If a project lasts weeks, refresh the handover each time you start over, and you will have a clean record of the work as a bonus.