Is your ChatGPT conversation too long? You'll usually notice in one of two ways. Either ChatGPT starts forgetting things you settled near the start (a name, a constraint, a decision), or it stops letting you add messages at all. The message people report seeing is: "You've reached the maximum length for this conversation, but you can keep talking by starting a new chat."
Both have the same root cause, and both have fixes. This guide explains what's going on, then goes through the fixes OpenAI supports, from quickest to most thorough. At the end we cover a different way to organize long work so a single chat never has to hold all of it.
A model can only read a limited amount of text at once, measured in tokens (in English, a token is roughly three quarters of a word). That limit is the context window. Everything ChatGPT considers when it answers has to fit inside it.
OpenAI publishes the total window for each plan. As of October 8, 2026, ChatGPT's pricing page lists:
| Plan | Instant models | Reasoning models | Input, OpenAI's estimate (Instant) |
|---|---|---|---|
| Free | 27K tokens | Varies | About 12 pages of text |
| Go | 54K tokens | 256K tokens | About 40 pages |
| Plus | 54K tokens | 256K tokens | About 40 pages |
| Pro | 128K tokens | 400K tokens | About 250 pages |
These numbers change with new models, so check OpenAI's pricing page for your plan.
The window isn't all yours. OpenAI says the space for your input is smaller than the total, because the same window also holds system instructions (including tools and personality), your memories if memory is on, and the model's own work: reviewing information, reasoning and writing the reply.
So a conversation that's been going for weeks, with pasted documents, web searches and long answers, fills the window much faster than the message count suggests.
OpenAI doesn't publish exactly how ChatGPT decides what to keep once a conversation outgrows the window. In practice, the symptom is that early details stop being used: the model contradicts a decision from message five, or forgets a rule you set at the start.
Even within the window, more isn't always better. Researchers at Stanford and elsewhere found that language models use information at the beginning and end of a long input best, and do noticeably worse with information buried in the middle ("Lost in the Middle", Liu et al., 2023). A 200-message chat puts most of what matters in the middle.
The "maximum length" message is a different limit from the context window. OpenAI doesn't publish how long a single conversation can get. When you hit it, the chat stops accepting new messages and you need a new one. The fixes below are about carrying your work over well.
The best fix is to never let one chat carry everything. When a side question comes up, branch instead of asking it in the main thread.
On chatgpt.com, hover over a message, click More actions (⋯) and choose Branch in new chat. You get a new chat with everything up to that message, and the original stays as it was. The tangent's messages never land in your main thread, so the main thread stays shorter.
When a chat is getting heavy (or has already hit the limit), ask for a handover note, then start fresh:
If the old chat has already hit the hard limit and won't answer, write the brief yourself from the parts you need, or use fix 5.
Projects keep related chats, files and instructions together, and they're available on all free and paid plans. OpenAI says projects have built-in memory: chats can draw on the other chats and files in the same project.
For work you want kept separate from everything else, choose project-only memory. Chats in that project can reference each other but nothing outside it. Shared projects always use project-only memory.
Some things should apply to every chat: your role, your preferences, a style you always want. Put those in memory so they don't depend on one long thread.
ChatGPT's memory has two parts. Saved memories are details you've told it to remember ("Remember that I write in British English"). Reference chat history lets it draw on past chats. OpenAI is clear that chat history doesn't keep every detail, and recommends saved memories for anything you want ChatGPT to always keep in mind. You'll find both under Memory in Settings.
That's also why "continue our last chat" in a new chat often disappoints: it relies on chat-history recall, not on the full old conversation.
If you need the old chat's actual content, not a summary:
conversations.json. Larger exports may be split into numbered conversation files.The export holds all your chats, so if the file is too big, OpenAI suggests uploading the numbered files separately, splitting the file, or uploading only the conversations you need. Pulling the one conversation you need into its own text file works too.
On Go and Plus, the reasoning models list a 256K token window against 54K for Instant. For a long document review or a chat that keeps losing early details, picking a reasoning model in the model picker gives the conversation more room. It doesn't remove the hard length limit.
OpenAI now turns pastes of more than 10,000 characters into an attachment automatically, to keep big pastes from eating the context window. Do the same on purpose: upload documents as files (in a project, ideally) rather than dropping them into the thread again and again.
All of the fixes above work around one fact: a chat is a single line, and everything you've ever said in it travels along. Branching helps, but in ChatGPT a branch is just another chat in the list, frozen at the moment you split it.
On a canvas, long work becomes several short chats laid out side by side, and you decide what each chat reads. That's how Slashspace works:
To be straight about the limits: sending everything in full means the context window still applies. Connect enough big sources and a request can exceed the model's window and fail. The fix is to disconnect what that chat doesn't need or pick a model with a bigger window. A Document node's text over about 500 KB (very roughly 80,000 words) is left out, so split long books into chapters. A canvas doesn't make the window bigger. It makes it obvious what's in it.
You can run it on your ChatGPT plan: Slashspace drives the official Codex command-line tool (and Claude Code, for Claude plans) on your computer, signed in with your own account, and usage counts against your plan. Your own API keys and local models work too.
Slashspace is $89/year, starting with a 3-day trial (card required), or $249 one-time lifetime, on Mac, Windows and Linux.
See pricing · Start a 3-day trial
Each conversation has a maximum length, and long chats with pasted documents, searches and long answers reach it sooner. OpenAI doesn't publish the exact limit. When you see the message, start a new chat and carry your work over with a summary, a project or an exported file.
OpenAI doesn't publish a maximum conversation length. It does publish context windows, which limit how much the model can read at once: as of October 8, 2026, 27K tokens on Free and 54K on Go and Plus for Instant models, with larger windows for reasoning models and on Pro.
It gives you more room. Plus lists 54K tokens for Instant models against 27K on Free, and 256K for reasoning models. Very long conversations can still outgrow that, so the habits above still help.
The conversation has likely outgrown what the model can use well. Older details either fall outside the context window or sit in the middle of a very long input, where models are weakest. Restate key decisions, keep them in project instructions or saved memories, or start a new chat with a brief.
No. Saved memories hold short, lasting facts, and chat-history recall doesn't keep every detail. For the substance of a long project, use a project with files, a handover brief, or an exported conversation.
Not directly, but you can bring its content along: paste a handover brief, move your work into a project so new chats can draw on it, or upload the exported conversation file as reference.