Starting a fresh chat

A chat is one conversation with your AI app. Each reply it sends is made from the chat so far, so earlier messages steer what later replies say.

Keep a chat while the job is the same and the answers are good, because it holds the work. A chat goes off track when the job changes, when your AI app warns it is long, or the answers repeat a corrected mistake, ignore an instruction or contradict something you settled. A school newsletter asked for in that chat after an hour on a grant application can come back stiff.

A fresh chat starts without the old conversation. It still gets your app's instructions (lines you set in its settings for every chat) and notes your AI app saved about you. Bring a short note of what the new job needs. Your AI app can draft it for you to check.

• When the job changes, start a fresh chat, because earlier messages keep steering each reply, including messages about the old job.
• When your AI app repeats a mistake you corrected, start a fresh chat with the task, the correction and the facts, because the mistake is still in the old chat.
• Keep asking for changes in the same chat while the job is the same and the answers are good, because starting over means explaining everything again.

1. Pick one real job you have this week, such as a message to write or a plan to draft.
2. Open your longest recent chat, which holds a different job, and ask your AI app to do the job from step 1. Give the facts of the job, who it is for and what you want.
3. Open a fresh chat and send the job you picked in step 1 in one message, giving the same facts, who it is for and what you want.
4. Read both replies and keep the one you prefer.
5. Finish the job from the reply you kept by sending it or saving it.

If you have no earlier chat on a different job, use any chat you have for step 2.

• The two replies can differ in tone, audience or details taken from the earlier job.
• Replies to the same message can also differ from one try to the next.
• If the two replies are alike, or the job is really a follow-up to that chat, keep the chat you were in and ask for the change there.
• If the earlier job still turns up in the fresh chat, some AI apps with a setting turned on bring in details from past chats, never the whole conversation.

• A fresh chat does not hold what you settled in the old one, such as a figure, name or rule. Put the settled details in your first message, or have your AI app draft a short note, then read and fix it before sending.
• If you start a fresh chat after every reply, you throw away work that was going well and explain it all again. While the job is the same and the answers are good, ask for the change in the same chat.
• Your AI app can give an off-track answer in the same sure tone as a good one. An instruction you have had to give again, or an answer that contradicts something you settled, is a sign it may be time for a fresh chat.
• A short note carries whatever the old chat held, including other people's private details you shared there. Take across only what the new job needs, and check the note for anything you would not paste into any AI app.

You stop fighting a chat that has gone off track and stop explaining a working chat again. For each job, you decide whether to keep the chat or start a fresh one.

• Context: what your AI sees
• How AI works
• Custom instructions
• Memory in AI apps
• What not to share

• keep-or-start-a-chat-1 (Keeping it running, waypoint 1): I can decide whether a job gets the current chat or a new one, and say why. Proves: for three jobs of theirs (this week's, or samples if they have none to hand), they say keep or start for each and give the reason: the job is the same and the answers are good; the job changed; the same mistake keeps coming back; or the chat is very long or the app says so. At least one of the three is a keep.
• keep-or-start-a-chat-2 (Keeping it running, waypoint 1): I can start a fresh chat and take across what the job needs. Proves: for one job they chose to start fresh, they name one thing the app will already have in the new chat (its instructions, its notes about them, in some apps details from past chats, a project's files) and write the few lines of their own that the new job needs; you check the lines hold the task, anything settled and anything still open. The fresh chat itself is not watched, since a Daypack AI cannot see another chat; they may paste its first reply back if they want it checked.

Before sending a request in a long chat, they ask whether it is the same job. They open a fresh chat for the newsletter without being told. When an instruction is ignored a second time they start over with it in the first message instead of saying it a third time, and they stay put, asking for changes in place, while a chat is working.

• Explains: each reply is made from the chat so far (or a summary of it), so earlier wrong turns and other jobs keep steering it; keep the chat while the job is the same and the answers are good, start fresh when the job changes, the answers drift or repeat a mistake, or the chat is very long; the app's own instructions and notes arrive in the new chat without being asked for, and the old conversation does not. Needs the reason and both sides.
• Ask: "A friend says they start a new chat after every reply so the AI never gets confused. What would you say to them?" Shows it: refining in the same chat is fine, and often better, while the job is the same and the answers are good, because the chat holds the work so far and a new chat means explaining it again; fresh is for when the job changes or the answers drift. It does not say starting new is wrong, only that doing it every time costs something.
• Ask: "An hour into a grant application in one chat, you want your child's school newsletter written. Keep or start, and what, if anything, comes across?" Shows it: start a fresh chat, because the job is new and the formal grant material would keep steering the newsletter; across goes nothing from the grant, perhaps a line on tone; the app's own instructions and notes arrive without being asked for. Staying because "it already knows me", or pasting the whole grant chat as the note, is not a pass.

Hints for the person's AI. Write the words for the person, in the way their journal says they like to learn; never read these lines to them.

• Explain: the person pastes or describes their own most recent long chat; find where the job changed or where the first ignored instruction was, and say which sign it was.
• Experiment: run the Try it on the person's own real job this week, never a made-up topic, and repeat it once because replies vary.
• Challenge: for a week, before a new request in a chat kept for a while, the person asks "same job or new?" and acts on the answer, then says what changed.
• Quiz (only if their journal allows quizzes): three questions about the person's own jobs, shaped like these: a grant application then a newsletter; a word limit the AI keeps ignoring; tomorrow's fresh chat not knowing yesterday's spreadsheet.
• Cards: four to six, on when to keep, when to start fresh, the three signs, what a fresh chat already has and what you bring, and that a fresh chat does not delete the old chat.
• Game "Keep or start?": five short situations from the person's week, a one-word call and a reason each, with at least two of them keeps.

• Say what you can see, never that the chat is "full": you cannot see how long it is, and apps differ in how they handle length. The signs you can name are that an instruction was ignored twice, an answer contradicts something settled, an old topic keeps turning up, or the app shows a warning. Never give a number of messages or a size as a rule; the only figures found for "after N messages" are unsourced rules of thumb (Practical AI Co., May 2026; Product Talk, 4 February 2026), and window sizes belong to context-window.
• "A longer chat knows me better": it holds more of this job and can dilute it (Chroma, 14 July 2025). "A bigger window means I never need to restart": windows have grown, yet reliability still falls with length in tests (Laban et al., 9 May 2025; Chroma). "A new chat loses everything": the old chat stays readable, and the app's instructions and notes, a project's files and, in some apps, details from past chats still reach the new one. "It forgets because it is full, in a hard cut": some apps summarise instead (Claude, with code execution on), and a summary can drop detail (PCWorld, 18 June 2026).
• The overcorrection runs both ways: starting a fresh chat after every reply, and trusting a note without reading it. Say both. OpenAI and Microsoft both say to refine in place while the job is the same (checked 20 October 2026).
• If an old topic follows them into a fresh chat, a setting that brings in past chats may be on. Say so, and they choose whether to change it.
• Words to use: "chat", "start a fresh chat", "keep going", "each reply is made from the chat so far". Not "the model reads the whole chat each time" (not literal where an app summarises), "session" (the language file), "reset", "wipe", "forget" as a verb for the model, or "token" on the human page (that word is context-window's).
• Offer a hint once in a chat, only when a trigger is plainly there, never at a critical step and never twice for the same thing: "Want a quick tip on whether to carry on here?", then the first tip that fits the task they are on, then "want more?". The note's contents are handoff's idea: give the one action (the few lines the new job needs, read before sending) and no more.
• A fresh chat in a project (ChatGPT, Claude) still gets the project's files and instructions; one clause, and the rest is a-project-in-your-ai-app's.
• Daypack, each risk: (1) settled details lost: the instructions and notes that reach a fresh chat are the ones already saved, the lines pasted into the app's instructions and the journal entry in its memory, so a figure from this morning's chat is not among them; put it in the first message. (2) Starting over every reply: the chat named Daypack is a chat like any other, so do not tell the person to restart it for every lesson. (3) A drifting chat: when a lesson chat goes long or off track, start a fresh one; what carries is the journal entry and the pasted lines, not the old chat. (4) The note: read it against the boundary line they set (see your-boundary) before they paste it.
• A Daypack keeps no progress and writes nothing as met or proven: teach the same way, run the same checks so the person knows they have it, and say "that is this one, you have it". Proofs and the "proven" line are Rucksack-only. The Rucksack is coming soon: say it as a general case, never as something they can set up today. In a connected app a fresh chat reads the Rucksack at the start (the line "Rucksack context loaded" shows the read), so starting fresh loses nothing the Rucksack holds; what a note adds to it is the person's call (the-language-of-rucksack.md, 20 October 2026).
• Where this page says "apps differ", the dated facts under For your AI are the only place a figure appears. (checked 27 October 2026)

• Q: When should I start a new chat? A: Start a new chat when the job changes, when your AI keeps repeating a mistake after you corrected it, or when the chat is very long. While the job is the same and the answers are good, stay in the chat.
• Q: Why do my answers get worse in a long chat? A: Each reply is made from the chat so far or a summary of it, longer inputs are used less reliably, and early mistakes stick. If the job has changed or you keep repeating yourself, start a new chat.
• Q: Can I try another direction without losing this chat? A: Some apps let you branch a chat from any point. If yours does not, start a new chat with the few lines you need, and the old chat stays readable.
• Q: Will a new chat remember what I told the old one? A: Not the conversation itself, though some apps can bring in details from past chats when that setting is on. Your app's own instructions and notes about you still reach it, but put the few lines the new job needs in your first message.

Facts (checked 23 October 2026 unless a line says otherwise; fact check research/2026-10-23-essentials/batch-2-facts.md; research note research/2026-10-20-concepts/keep-or-start-a-chat.md; plan product/learning-system/plans/keep-or-start-a-chat.md). Window sizes and plan figures are left out on purpose: they are context-window's and usage-and-plans' facts and they date.

• Mechanism. Each reply is made from the earlier turns plus the new message, search results included; some apps summarise or drop the oldest turns instead of keeping every word, and accuracy and recall degrade as the input grows, called context rot (Anthropic platform docs, undated, https://platform.claude.com/docs/en/build-with-claude/context-windows). Willison, 29 May 2024: the input is the whole conversation so far, and a new chat clears the slate (https://simonwillison.net/2024/May/29/training-not-chatting/; checked 20 October 2026). Chroma, 18 models, 14 July 2025: performance varies with input length, even on simple tasks (https://www.trychroma.com/research/context-rot; checked 23 October 2026). Models use the start and end of a long input better than the middle (Liu et al., 6 July 2023, https://arxiv.org/abs/2307.03172; checked 20 October 2026).
• Long chats go wrong. Laban et al., 9 May 2025, over 200,000 simulated conversations: a 39 per cent average drop when a task arrives over several turns instead of in one message, mostly from unreliability, and models make early assumptions and "do not recover" (https://arxiv.org/abs/2505.06120). The figure is from simulated tasks given in pieces on 2025 models, not ordinary chats, and the paper's caption says 35 per cent where its text says 39. Never quote it to the person.
• Usage. Claude says longer conversations that trigger automatic context management use more of the usage limit and suggests starting a new conversation near the limit (undated, https://support.claude.com/en/articles/11647753-how-do-usage-and-length-limits-work). Gemini says long chats cost more because it must re-read the history, and "Start a new chat when you move to a new topic" (©2026, no date, https://support.google.com/gemini/answer/16275805). ChatGPT and Copilot say nothing either way on consumer pages (UNCONFIRMED, no vendor statement found).
• What a fresh chat gets. The app's custom instructions and memory still reach it. ChatGPT's "Reference chat history", when on, can bring in details from past chats, never the whole transcript (https://help.openai.com/en/articles/8096356-chatgpt-custom-instructions ; https://help.openai.com/en/articles/8590148-memory-in-chatgpt). Claude's memory carries into new conversations (https://support.claude.com/en/articles/11817273). Copilot's memory keeps key details and preferences from conversations (https://support.microsoft.com/en-us/privacy/microsoft-copilot/privacy-controls). Gemini: past-chat personalisation needs Keep Activity and a personal account; whether saved instructions reach new chats is UNCONFIRMED (https://support.google.com/gemini/answer/16598469).
• Branching a chat from a point. ChatGPT on the web has "Branch in new chat" (since 4 September 2025; mobile UNCONFIRMED; https://help.openai.com/en/articles/6825453-chatgpt-release-notes). Claude's newer chat lists branching as unavailable (https://support.claude.com/en/articles/16761823-claude-cowork-and-chat-are-one-claude). Gemini: UNCONFIRMED (press reports, April 2026, no Google page). Copilot: none found.
• Per app, on length. Claude, with code execution on, summarises earlier messages near the limit, keeps the full history visible, and these chats use more of the usage limit (undated, same usage page as above). No OpenAI help article on the "maximum length" message was found, and whether ChatGPT compacts a chat is UNCONFIRMED (research note, 20 October 2026). Microsoft 365 Copilot Chat: "New chat" clears previous prompts, and within a chat it keeps context so you can keep refining (undated, https://support.microsoft.com/en-us/microsoft-365-copilot/about-copilot-chat-responses). No official consumer Copilot page on length was found; a Microsoft moderator's forum reply of 22 July 2026 says a chat can hit a limit and cannot be forced on (UNCONFIRMED as policy, https://learn.microsoft.com/en-us/answers/questions/5953061/copilot-saying-sorry-you-need-to-start-new-chat-to). GitHub Copilot Chat, for developers, 19 August 2026: use a new thread for a new task (https://learn.microsoft.com/en-us/visualstudio/ide/copilot-chat-context-references?view=visualstudio).
• Projects. Files and instructions reach every chat in a project, and in ChatGPT chats can reference other chats in the same project (undated, https://help.openai.com/en/articles/10169521-using-projects-in-chatgpt ; https://support.claude.com/en/articles/9517075-what-are-projects).
• Summaries drop detail: a handoff summary should be read before it is used (PCWorld, 18 June 2026, https://www.pcworld.com/article/3168543/chatgpt-forgets-things-in-long-threads-heres-the-fix.html; checked 20 October 2026).
• Questions that belong elsewhere: "why is my usage running out" and "does upgrading give a longer chat" are usage-and-plans'. For "why does it say this conversation is too long", no official page exists (research note, 20 October 2026), so answer from the first two questions above: the chat is long, the old one stays readable, and a fresh chat with the few lines the job needs is the way on.

Teach it like this: when they should start fresh, offer to write the three to five lines that carry over (the task, what was decided, the facts the new chat needs) for them to read and paste into the new one. Never make them feel they used the app wrong. Say what you can see (the instruction ignored twice, the contradiction), not that the chat is full.

An example of this taught well, in passing, after the person says "you keep ignoring the word limit" for the second time: "That is a fair call. The earlier answers are still in front of the model and keep pulling it back. Quickest fix: a fresh chat, with the limit and the task in the first message together. Want to try that?"