ChatGPT Conversation Too Long? How to Continue in a New Chat

You are in the middle of something that finally works. Then ChatGPT stops and tells you that you have reached the maximum length for this conversation, but you can keep talking by starting a new chat.

Starting a new chat is easy. Keeping the work is the hard part.

This is not a bug, and it is not a problem with your account. It is also not something OpenAI publishes a number for, which is why so many people assume they did something wrong.

Here is what is actually happening, what to capture before you hit the wall, and how to restart without going back to the beginning.

Why ChatGPT says your conversation is too long

Every model has a fixed amount of text it can hold at once. Your entire thread counts toward it: your messages, every reply, any files you uploaded, and anything the model wrote back. When the thread fills that space, the conversation stops accepting new messages.

Two things follow from this that are worth knowing.

The first is that quality drops before the wall, not at it. Long before you see the message, the model is working with a crowded context, and it starts leaning on the most recent exchanges instead of the decisions you made hours ago. If you have noticed answers getting vaguer near the end of a long thread, that is this.

The second is that a new chat is genuinely empty. It does not inherit your thread. Memory features and Projects help with some of this, but neither one reliably carries the specific state of one long working session into the next.

What to capture before you hit the limit

The best moment to do this is when the thread starts feeling heavy, not when it stops.

Ask the current chat to write you a handoff. Not a summary of the conversation, which will read like meeting minutes and drop the things you need. Ask for these six parts by name:

Goal. What you are actually trying to produce, in one or two sentences.

Decisions. What has been settled and why. The why matters, because without it the new chat will reopen questions you already closed.

Constraints. Rules, formats, tone, things to avoid. This is the part people forget, and it is the part that makes the new chat feel wrong even when the facts are right.

Current state. Where the work stands right now, including the latest version of whatever you are building.

Open questions. What is still unresolved.

Next step. The single next action, stated concretely.

Copy that handoff somewhere outside ChatGPT before you start the new chat. You will want to edit it, and you cannot edit it once the old thread is closed to you.

If you already hit the limit

You can still scroll the old thread. It is closed to new messages, not deleted.

Open it, find the last stretch where the work was going well, and copy the parts that carry state: the current version of the output, the rules you established, the decisions you argued through. Then write the six parts yourself from that material. It takes a few minutes and it is more reliable than a summary the model writes under pressure at the end.

Make the new chat prove it understood

This is the step almost everyone skips, and it is the one that catches problems while they are still cheap to fix.

Paste your handoff, then before asking for any work, ask the new chat to state back what it thinks the goal is, what constraints it is operating under, and what it believes the next step is.

You will find the gaps immediately. Usually it is a constraint that got compressed away, or a decision it has quietly reversed. Correct it in one message, then start working.

Whole chat, summary, or selected messages

The instinct is to paste everything. It is usually the wrong call.

More text is not more context. When you paste a long transcript, the important decisions sit in the same pile as the false starts, the corrections, and the tangents you abandoned. The model has to work out which parts still apply, and it does not always work it out the way you would.

A rough guide:

Paste the whole thread when it is short, when the reasoning matters more than the conclusion, or when you genuinely do not know yet which parts are load bearing.

Paste a summary when the thread was long and exploratory and only the conclusions survived. Read the summary before you use it. Summaries drop constraints more often than they drop facts.

Paste selected messages when you know exactly which exchanges hold the state. This is the most accurate option and the most work, and for ongoing projects it is usually worth it.

If you are switching to a different assistant rather than a new ChatGPT thread, bring more of the constraints than you think you need. Tone, format, and output style do not carry across models the way facts do.

The manual options, honestly

People solve this with summaries, exported PDFs, Markdown files, Projects, instruction documents, runbooks, and notes in Notion or Obsidian. These work. They are editable, you can read them, you control them, and they cost nothing.

What they cost is upkeep. You maintain them, you copy things in and out, and when you move between assistants you do the search and the carrying yourself. If you do this across assistants regularly, there is a broader guide to carrying your context between assistants.

If this happens to you every week

Unibase Memory is a Chrome extension that keeps your conversations from ChatGPT, Claude, Gemini, Grok, Kimi and DeepSeek in one searchable place, along with web pages you save. When you need to restart, you search for the relevant conversation, select the specific messages that hold your decisions and constraints, and send them into the input box of whichever assistant you are using. You edit them there before you submit.

It does not extend ChatGPT's limit and it does not decide for you what matters. It removes the part where you go looking for where you said it.

Unibase Memory: https://chromewebstore.google.com/detail/unibase-memory-%E2%80%93-private/edmncknbiihfoakimejbepnaeemaaamf