# "Copilot Conversation Limit: Why It Says Start a New Topic"

> "Copilot's limit is a turn cap, not a token ceiling, which means you can plan for it. Here is how to budget your turns and tell the five different limits apart."

Published: 2026-08-14 by folk team

You are mid-task, and Copilot stops answering the way it was: **the conversation has hit its limit and you are being pointed at the New topic button next to the input field.**

Here is the thing worth knowing before you do anything else. Copilot's limit is not the same kind of limit as every other assistant's. ChatGPT, Claude, Gemini, and DeepSeek all fail on an invisible token budget you cannot see or count. Copilot fails on a countable number of turns. That is a far better problem to have, because you can plan around a number you can count.

> **Quick answer:** Copy anything important out of the thread, click New topic, and paste a short summary of what you were working on as the first message. Copilot caps conversations by number of exchanges, not by length, so the fix is instant. The better move is to write that summary a few turns before the cap rather than after it.

## Which Copilot limit did you actually hit?

"My Copilot limit changed" is almost always one of the other four limits, not the conversation cap. They look similar and they clear in completely different ways.

| Limit | What it looks like | Scope | What clears it |
|---|---|---|---|
| Conversation turn cap | Copilot asks you to start a new topic | This one thread | Starting a new topic, immediately |
| Daily or rolling usage cap | You are out of uses for now, in every conversation | Your account | Time |
| Model or capability cap | Only the advanced model, image generation, file analysis, or voice is blocked, while ordinary chat still answers | That one capability | Time, or falling back to the standard model |
| Work or school policy | Your organization restricts what Copilot can do, see, or connect to | Your tenant | Your IT administrator, not you |
| Temporary throttling | Refusals or degraded answers during periods of heavy demand | Everyone | Time |

Only the first row is fixed by starting a new topic. If you start a new topic and hit the same wall on your first message, you are in one of the other four and this guide's fix will not help you.

## Fix it: carry the thread across the break

The cap is per conversation, so the recovery is quick. The only thing that makes it painful is losing what the old thread knew.

1. **Do not click New topic yet.** You still have the thread, which means you still have one chance to get a summary out of it.
2. **Ask for a handoff summary in the old thread.** Something like: what are we trying to accomplish, what have we decided, what would you get wrong by guessing, and what is the very next step.
3. **Save it outside the chat.** A file or a note, not your clipboard. Clipboards get overwritten by the next thing you copy.
4. **Copy the exact values separately.** Numbers, names, file paths, code. Summaries paraphrase, and paraphrasing quietly corrupts precisely the things you cannot afford to have wrong.
5. **Click New topic** next to the input field.
6. **Paste the summary as your first message**, followed by: "Continue from here. Confirm you have it before doing anything."

That is the whole recovery. The rest of this guide is about not needing it.

## Budget your turns, because you can count them

This is the part that does not apply to any other assistant. The University of Helsinki's IT helpdesk [documents the cap as a maximum of 30 answers per conversation](https://helpdesk.it.helsinki.fi/en/instructions/information-security-and-cloud-services/cloud-services/microsoft-copilot-university), after which a new topic must be started. Whether that exact number holds for your account today or not, the shape of the limit is the useful part. It is a count, and counts can be spent deliberately.

Four ways to spend them well:

- **Front-load context into fewer, richer prompts.** Three clarifying exchanges plus one real question costs four turns. The same request written once, with the constraints, the format you want, and the background included, costs one. On a token budget that trade is roughly neutral. On a turn cap it is a straight saving of three.
- **Do not use turns for acknowledgements.** "Thanks, that works" and "ok, continue" cost the same as a real question. In a capped conversation they are pure overhead.
- **Write the handoff summary on a schedule, not on an error.** If you are planning against a cap of thirty, write it around turn twenty five, while there is still room for the model to think properly. A summary produced at the ceiling is written under the worst conditions and reads like it.
- **Let Edge carry the context instead of your thread.** Copilot in Edge can read the page you are on, so pointing it at the page beats pasting the page into the conversation. Anything you do not paste is a turn you did not spend.

There is a second, less obvious reason to use the button. **Start a new topic when you actually change topics.** Carrying an unrelated earlier discussion into a new question does not just cost you turns, it degrades the answer, because the model is balancing two contexts at once. The New topic button is a quality tool as much as a reset.

## Signed out versus signed in

If your conversations end far sooner than expected, check whether you are signed in. Reporting on this is third-party rather than official, so hold the numbers loosely: a [2024 write-up on Copilot tips](https://vladtalkstech.com/microsoft-copilot/5-microsoft-copilot-tips-and-tricks-to-get-the-best-results/) documents 5 prompts per conversation when signed out, rising to up to 30 prompts per conversation when signed in with a Microsoft account, along with saved history and personalization.

Two cautions. That figure is from 2024 and describes a consumer product Microsoft revises frequently, so treat it as reported rather than as current official documentation. And if you are on a work or school account, your organization's policy sits on top of all of this and can be more restrictive than anything published publicly.

## Why any of this exists

Underneath the turn cap is the same constraint every assistant works inside. On each turn the model is re-sent the whole conversation: the system prompt, every previous message, every attachment, and space reserved for the reply. That total competes for one fixed budget measured in [tokens](/glossary/token-limit), each roughly three quarters of a word, inside a [context window](/glossary/context-window) the model cannot exceed.

A turn cap is a blunter enforcement of that reality than a token ceiling, and it cuts both ways. It can end a conversation of thirty short questions that was nowhere near full. It also gives you something no token budget does, which is a number you can see coming.

Where some assistants respond to a full thread by [compacting](/glossary/compaction) it, condensing the older turns into a summary and continuing, Copilot's answer is a button. Nothing carries across it except what you carry.

And the cap is not the only way a Copilot conversation degrades. Well before any limit, models weight recent text more heavily than distant text, so a formatting rule or constraint you set early gets technically retained and effectively ignored. That is [context rot](/glossary/context-rot). If Copilot starts violating something you established at the top of the thread, re-state it in your next message rather than pointing back at the old one.

## Stop having this problem

Budgeting turns is a genuinely useful skill and it is also, plainly, unpaid administration. You are tracking a counter on the assistant's behalf, remembering to write the summary, storing it, and pasting it back in, and if you miss the moment you retype from memory.

The reason the workaround feels like a chore is structural. Your context lives inside a thread, inside one app, on one device. The thread is the unit of memory, so it ends when the thread ends, and it never reaches your phone. The popular third-party fixes for this are browser extensions, which is a reasonable answer at a desk and no answer at all anywhere else.

[folk](/) inverts that. Its memory is not scoped to a thread, so there is no turn counter to run down and no topic to start over. It reaches you in iMessage and Telegram, which means what you worked out on Tuesday is still there on Friday, from your phone, without a summary in between.

If you already have context worth keeping elsewhere, you do not have to retype it. You can [import your ChatGPT and Claude memory](/import-memory) in one paste.

Hitting a wall on another assistant? Those fail on a token budget instead of a turn count, so the fixes differ: see the guides for [ChatGPT](/blog/chatgpt-maximum-length-conversation-fix), where you can branch out of a full thread, and [Gemini](/blog/gemini-forgetting-context-fix), which tends to forget quietly rather than stop. The overview is [why every AI assistant forgets your conversation](/blog/why-ai-assistants-forget-your-conversation).

---

Canonical page: https://www.folk.com/blog/copilot-conversation-limit-new-topic
More about folk (for AI agents): https://www.folk.com/llms.txt · full context: https://www.folk.com/llms-full.txt
folk is a personal AI that lives in your texts (iMessage, Telegram, WhatsApp). Pro $20/mo, Max $100/mo. Made by Nozomio Labs. Sign up: https://www.folk.com/