Why ChatGPT gets worse in a long chat and forgets your instructions

In the morning you told the chat to keep answers short and skip the bullet points. For the first few messages it did. By the afternoon it is writing long bulleted lists again and mixing up two proposals you discussed earlier. This is what it looks like when ChatGPT forgets in a long conversation, and the same happens in Claude.

It isn't a bug, and nobody is doing it on purpose. It is how language models work, and a few simple habits get around it.

Why ChatGPT forgets what you agreed on

A model doesn't remember a conversation the way a person does. For every answer it looks at the text that fits into its context window. Anthropic's documentation calls this the model's working memory. It holds everything the model can refer to while writing a reply.

The window is large, but it has an end. There is also a less obvious problem. According to Anthropic's documentation, as the amount of text in context grows, the model's accuracy and recall drop. More context is not automatically better.

In practice, your morning instruction sits far back, buried under pasted emails and one correction after another. It gets less attention than whatever you wrote a minute ago. In a very long chat the oldest parts can be summarised or drop out. In some setups Claude summarises earlier messages on its own when a conversation gets close to the limit.

That is where the drift comes from. The style you fixed hours ago comes back, and answers blend two topics from the same chat.

One task, one chat

This is the simplest habit. The client proposal goes in one chat, the order spreadsheet in another. When everything is in one window, the model sees both topics at once and starts mixing them.

In Claude there is a second benefit. Conversation length affects how fast you use your plan's limit, and long chats that Claude has to summarise use more of it. Anthropic itself suggests starting a new conversation in that case.

The same goes for documents. Instead of pasting your whole store terms and asking about returns, paste only the returns section and say that's the part you mean. With a long file, point to the exact place, a section number or a heading. Less text in the window leaves more attention for what matters.

When the answers start to drift

You don't have to start from scratch. When you notice the drift, write something like this in the same chat.

Summarise briefly what we've agreed. The goal, the decisions, the data we're using and the writing rules. No commentary.

Read the summary and fix anything that came out wrong. Then open a new chat and paste it in as the first message. The model starts with a clean slate, but with the things that actually matter.

It's better to do this early than to correct the same thing again and again in the old chat. Every correction adds more text to the window, and the model still has to read all of it.

Keep permanent instructions out of the chat

If you type the same things into every chat, like what the business does and the tone you use with clients, move them into the settings.

ChatGPT has two places for this. Custom instructions, under Personalization in Settings, apply to all your chats. Projects keep the chats, files and instructions for one topic together. Project instructions work only inside that project and override your general custom instructions.

Claude works much the same way. Instructions for Claude in Settings apply to every conversation. Projects have their own files and instructions that apply only to chats in that project. On paid plans, when a project holds a lot of files, Claude loads only the parts that are relevant to the question.

Anthropic recommends keeping project instructions short and about the general context, and leaving today's task for the chat itself. That rule works for ChatGPT too.

If someone in your business pastes the same data into a chat every day, put those hours into the manual work calculator. Sometimes that's a sign the task belongs in an automation rather than in another chat.

When the answers start to drift, don't keep correcting them. Ask for a short summary and start a new chat.