Workbranch › Guides

Why ChatGPT forgets earlier parts of a long conversation

Two hours into a chat, ChatGPT contradicts a decision you made at the start or asks for something you already gave it. That isn't memory failing; it's the conversation outgrowing what the model can hold. Here's how much fits on each plan and how to move on cleanly.

Updated 20 October 2026

Checked against OpenAI's pricing page and help pages on 20 October 2026. Window sizes change with new models, so treat the numbers as of October 2026.

How much a chat can hold, by plan

OpenAI lists context windows on its pricing page, in tokens (pieces of words) and in rough pages of text you can put in:

Two things on that page matter more than the totals. First, the window is shared: system instructions (including tools and personality), your memories if Memory is on, and ChatGPT's own reasoning and response generation all use part of it. Second, the space left for you is an estimate that shifts with the features in use. So a chat can run out of room sooner than the page count suggests, especially if you've pasted long documents or the model has been thinking hard.

Instant is the fast setting; the reasoning figures apply when you choose a thinking level. On Go, Plus and Pro, the reasoning window is several times larger, so a thinking level is the better choice for work with a lot of material.

Memory isn't the same as the chat

People often expect Memory to fix this. It works differently. Within one chat, ChatGPT uses the conversation itself as context. Memory, under Settings > Personalization > Memory, carries relevant preferences and details across chats, and OpenAI says it doesn't keep every detail from every conversation. ChatGPT decides what's relevant for each answer. So Memory may know you prefer metric units, but it won't hold the twelve requirements you agreed on in message four. For managing it, see how to manage ChatGPT memory.

Carry a long chat into a new one

1. In the long chat, ask for a handover summary. For example: "Summarize this conversation for a fresh chat: the goal, every decision we made, the constraints, open questions and next steps. Keep it under one page."
2. Check it against the chat. The summary is written by the same model that's losing track, so scroll back and add anything important it missed.
3. Select New chat and paste the summary as your first message, with a line such as "Continue from here."
4. Re-attach only the files you still need. Large inputs use the window too.
5. Keep the old chat. It stays in your history, so you can go back to the full discussion.

Use a project for context that has to last

For work that spans many chats, a project is a better home than one endless thread. Projects keep related chats, files and instructions together, and ChatGPT uses that context in every chat inside the project.

1. Select New project in the sidebar, then name it.
2. Add reference files, and put standing rules under Project settings as project instructions.
3. When a chat produces a decision note or summary you'll need again, open the message menu on that reply and save it to the project's sources (OpenAI says the label may vary).
4. Start each new subtopic as a new chat in the project instead of continuing the old one.

With project memory, chats can also draw on other conversations in the same project. On Free a project holds 5 files; on Go and Plus, 25; on Pro, Business, Enterprise and Edu, 40. Setup is covered in how to set up ChatGPT Projects.

Good to know

• Long chats can use more of your allowance. Where an allowance applies, OpenAI says usage depends partly on how much context is processed, so very long threads can bring you to a limit sooner. See what to do when you reach a ChatGPT limit.
• ChatGPT Work has a much larger window. OpenAI lists a default 500,000-token context window for GPT-6.1 Sol in ChatGPT Work and Codex, with plan-dependent access and its own allowances.
• Long chats load faster than they used to. On the web, ChatGPT now loads long conversations in smaller sections. That helps speed, not how much the model can take in.

If it doesn't work

• The new chat still gets things wrong. The summary probably left something out. Add the missing decision explicitly, as a list.
• ChatGPT ignores a long document you pasted. Split it, or ask about one section at a time, and say which part matters.
• Answers bring up an old, unrelated preference. That's Memory, not the chat. Correct it or delete it in Memory settings.
• A project chat brings in another chat's details. Chats in the same project can draw on each other. If one shouldn't influence the rest, delete it or, where the project's memory settings allow, move it to another project.

Keep decisions outside the chat

The underlying problem is that decisions live inside a scrolling thread. Workbranch's free chat organizer takes a ChatGPT share link and sorts the conversation into topics with tasks, decisions, open questions and ideas, which also makes a good handover for your next chat. It needs no account and deletes the result after 24 hours. For more on why long threads drift, see how follow-up questions derail AI chats.

Sources

• ChatGPT plans and pricing (OpenAI, checked 20 October 2026)
• Memory in ChatGPT (OpenAI Help Center, checked 20 October 2026)
• Projects in ChatGPT (OpenAI Help Center, checked 20 October 2026)
• ChatGPT release notes (OpenAI Help Center, checked 20 October 2026)
• What is ChatGPT Go? (OpenAI Help Center, checked 20 October 2026)
• What is ChatGPT: FAQ (OpenAI Help Center, checked 20 October 2026)
• ChatGPT Work and Codex (OpenAI Help Center, checked 20 October 2026)
• About ChatGPT Pro tiers (OpenAI Help Center, checked 20 October 2026)

Frequently asked questions