Which tool to choose
All assistants suffer from conversations that are too long, some sooner some later, because each has a limit on how much text it can keep in focus at once. If you do long sessions on a project, an assistant that lets you save a fixed context (a project, persistent instructions, a memory) is preferable, so the fixed points don't depend on the chat scrolling. In any case the defense that works everywhere is the same: a periodic summary and a document of yours with the things not to lose.
How to do it
The more a conversation grows, the harder the AI finds it to keep everything present: the old parts fade and errors you'd already corrected reappear, or contradictions with what was decided earlier. The breaking point doesn't announce itself: you notice it when the answers get worse. Better to prevent with a few habits.
- Keep a document of your own with the fixed points: decisions made, constraints, names, style choices. It's your memory, not the AI's.
- Every time you decide something important, update it (even just by copying the key sentence).
- When you feel the chat getting heavy or the answers declining, ask the AI for a concise summary of what you've established.
- Open a new chat and paste the summary as the starting point: restart light, with the facts, without the weight of the history.
- For very long tasks, break them into several chats from the start, one per phase, connected by your context document.
The operational syntax for the handover summary:
We've been working for a while. Before continuing, give me a concise summary of:
- the decisions we've made
- the constraints and preferences I need to remind you of
- where we are and what's left
Write it so that I can paste it into a new chat and you pick up exactly from here.
Save that summary in your document and use it to restart. The document you own is the only memory that doesn't fade with the chat's length and that you can also carry over to another assistant.
A concrete example
Andrea is writing a manual with the AI and after two hours of chat notices it has started using again a term they'd decided to avoid at the start of the session. The conversation is too long, the beginning has faded. Andrea asks for the summary of the fixed points (tone, glossary, decided structure), pastes it into a new chat and restarts. The AI becomes coherent again because it has the rules in front of it once more, fresh, instead of buried under hundreds of messages. The manual doesn't lose the thread.
When it does NOT work (and how to fix it)
If the AI contradicts something you'd decided earlier
It's the classic signal of the too-long chat: the decision has ended up too far back to stay in focus. Don't argue inside the same heavy chat: ask for the summary, restart in a new one by pasting it, and restate the contradicted decision at the top. In a fresh chat with the rules in front of it the error doesn't reappear.
If the summary it gives you is incomplete or wrong
The AI has already lost some pieces. Don't trust its summary blindly: compare it with your document of fixed points and correct it before using it to restart. That's exactly why you keep the document yourself: the AI summarizes, you verify.
If you don't want to lose the thread by opening a new chat
The fear is a misunderstanding of the method. You don't lose the thread: you transfer it. The summary plus your document contain everything needed to pick up; what you leave behind is only the history that was already slowing the AI down. The new chat isn't a starting over from scratch, it's a restarting lightened.
A tip from someone who actually uses it
Don't wait for the chat to collapse before acting. Get into the habit of asking for the summary at regular intervals, before the answers get worse, and of keeping it in your document. The reliable memory in a long job isn't the AI's, it's the file you own: update it as you'd update meeting notes, and no chat that's too long will be able to make you lose what matters.
Frequently asked questions
Why does the AI get worse when the chat is long, if it "remembers" everything?
Because there's a limit to how much text it can actually keep in focus at once. Beyond that limit the older parts weigh less or fall out of range, even if technically they're still there. It doesn't forget like a person: it's more like a desk that at a certain point is too cluttered to find the old papers on it.
Is a long chat or many short chats better?
For long jobs, many chats connected by your context document. A single, endless chat accumulates weight and falters; several short chats, each with a fresh summary at the top, stay sharp. The thread is held by the document you own, not by the length of the conversation.
Doesn't the assistant's automatic memory already solve the problem?
It helps, but it's not enough and trusting it entirely disappoints. Automatic memories keep some things and leave others, according to criteria you don't control; on a long, precise job you can't know what they'll keep. The context document you manage yourself is the only memory whose content you decide: for what really matters, keep it yourself.