ChatGPT keeps bringing up old conversations — how to start fresh

Fresh answers on the topic you asked — not last month's project bleeding in. It's not a glitch: ChatGPT references your past chats by design. Here's how to control it today, and why the controls feel blunter the more separate things you're running.

You ask ChatGPT something new and get something old back: a draft that arrives in last quarter's framing, an answer that quietly leans on a project you closed in March, a detail from one client surfacing while you're working on another. If that sounds familiar, the first thing to know is that nothing is broken.

This is a feature working as designed

ChatGPT carries context across conversations in two ways. Saved memories are discrete facts it stores about you — your role, your preferences, projects you've mentioned. And since April 2025, reference chat history lets it draw on your past conversations as a whole, so a new chat can pick up where an old one left off without you re-explaining anything.

Continuity is the point, and for a lot of use it's genuinely good — you stop repeating yourself. The friction starts when continuity crosses a line you didn't draw. The model decides which past context is "relevant" to the question you just asked, and its judgment isn't yours. Two clients whose situations rhyme, a personal question that pattern-matches to a work thread, an assumption from an old engagement applied to a new one — the same machinery that saves you re-typing is what drags yesterday into today. That's not a bug to report; it's a design tradeoff — continuity over separation — and you can move the slider. Here's how, lightest touch first.

Get one fresh answer right now: Temporary Chat

For a single question you want answered on its own terms, open a Temporary Chat. It doesn't use your saved memories or your chat history, doesn't save new memories, and won't appear in your history afterward. It's the closest thing ChatGPT has to a clean room: the answer is built from what you typed, nothing else.

Simply asking — "answer this fresh, don't draw on our previous conversations" — often helps too. But it's a request, not a boundary; the model can still be shaped by what it remembers. When fresh actually matters, use the Temporary Chat switch, not a polite instruction.

Remove the specific thing it keeps raising

If one detail keeps resurfacing — an old job title, a dead project, a client you've rolled off — go after that detail rather than the whole feature:

• Delete the saved memory. Open Settings → Personalization → Memory → Manage memories, find the entry, and remove it. You can also just tell ChatGPT in a chat to forget it.
• Delete the old conversation. OpenAI's published guidance says a deleted chat stops being used to inform future answers — so removing the March thread takes it out of what history-referencing can draw on. One catch: any saved memories ChatGPT created from that chat live separately and persist until you delete them from the memory list too.
• Prune on a schedule. A skim of the saved-memory list every few weeks keeps stale facts from steering new answers.

Turn off past-chat referencing entirely

For a fresh start everywhere, go to Settings → Personalization → Memory and toggle off Reference chat history. ChatGPT stops drawing on your past conversations, and OpenAI says the information it had gathered from your chat history is deleted from its systems within about 30 days. Reference saved memories is a separate toggle, so you can keep the useful facts-about-you while cutting the conversation carryover. (Turning saved memories off switches history-referencing off with it, but not the other way around.)

The tradeoff is that the switch is account-wide. You lose the carryover where it was annoying and where it was helping — there's no "reference my past chats about this topic, but not that one."

Keep workstreams separate: project-only memory

The closest ChatGPT comes to real separation is Projects with project-only memory. A project set up this way is a sealed context: its chats draw on the conversations and files inside that project, and nothing from your wider account — no saved memories, no unrelated history — comes through. For keeping one client's context out of another's, this is the best in-ChatGPT answer today.

Two caveats, as of September 2026: you choose the memory setting when you create the project or change it later under Project settings → Memory (OpenAI's 14 August 2026 release note; changes can take a few hours — see how ChatGPT project memory works), and the walls only contain what happens inside ChatGPT. The emails, documents, and meeting notes where the actual client context lives were never in any project to begin with.

Why the controls feel blunt

Here's the whole toolbox in one view:

Notice the shape: every control is a variation of all or nothing. That's because underneath, ChatGPT holds everything it knows about you in one pot, and decides for itself what to ladle into each answer. When it brings something old into something new, there's no per-answer trail showing which past detail it used or where that detail came from — you notice the bleed-through in the output, or you don't. ChatGPT's mid-2026 memory update — "Dreaming V3", rolling out from June 2026 — consolidated how it stores what it knows, and drew criticism for making that trail harder to see, not easier.

For one person with one stream of life, the pot is a fine design — continuity wins. But if you're running seven client engagements, aging parents' care, and a household, separation isn't a preference. It's the job. And a single opaque pot with account-wide switches is the wrong shape for it.

An assistant where you decide what it holds

This is the problem Equerry is built around, from the opposite starting point. Instead of remembering everything and deciding what's relevant for you, Equerry remembers only what you deliberately send it — and shows its work when it answers.

• You choose what it holds. Forward or CC an email thread, share a PDF or a screenshot, save a note, dictate a voice memo. That's the archive — nothing you didn't put there, so nothing you didn't put there can bleed in.
• Answers show their source. "What did the CFO say about pricing in October?" comes back showing where it came from when it draws on your notes, emails or the web.
• Nothing crosses over silently. If a detail from one client's thread surfaces while you're asking about another, the source chip shows you immediately which email or note it came from — and you can delete the fact behind it on a tap. That's the difference from the pot: not a promise that context never mixes, but a trail you can see and act on when it does.
• Everything remembered is visible and deletable. The full list of what Equerry has remembered is inspectable in one place, and any fact or source can be removed on a tap — a fresh start you can actually verify, not a toggle you take on faith.

None of this replaces ChatGPT for what it's good at — drafting, thinking out loud, general questions. Keep it, and use the controls above to manage the carryover. The shift Equerry offers is for the working context that was never in ChatGPT anyway: the threads, documents, and notes behind your clients, held by an assistant that remembers exactly what you gave it and briefs you from that — with receipts.

For the fuller picture of what ChatGPT keeps and doesn't, see does ChatGPT remember past conversations. If you've hit the other side of this coin — the saved-memory ceiling — here's what to do when ChatGPT memory is full.

Last updated 2026-09-08.

All product names, logos, and brands are property of their respective owners. Equerry is not affiliated with, endorsed by, or sponsored by any other product mentioned; comparisons are provided for informational purposes.