Why does ChatGPT forget previous conversations?

Yesterday's chat walked through the whole engagement. Today's opens blank, asking who the client is. Here's why that happens, the settings to check in order, the fixes that actually hold — and the one limit no setting changes.

The short version: ChatGPT forgets previous conversations because every new chat starts from an empty working thread. Continuity comes from a separate memory layer — a set of saved facts about you, plus best-effort references to your past chats — and that layer is selective, sometimes switched off, and capped on the older system. When the layer misses, the new chat genuinely starts from nothing. That's not a bug; it's how the product is built. What follows is why, what to check, and how to make it hold on to more.

A new chat starts empty by design

When you send a message, ChatGPT composes its answer from the conversation in front of it — the current thread — plus whatever its memory layer chooses to add. It does not go back and reread last Tuesday's conversation. If nothing from that conversation made it into the memory layer, the new chat has no way to know it ever happened.

The memory layer has two mechanisms, and knowing which is which explains most of the "why did it forget?" moments:

• Saved memories. Specific facts ChatGPT distills about you over time — your role, projects you've mentioned, preferences. Depending on your account, these appear as a list of entries or as a running memory summary under Settings → Personalization → Memory. Crucially, they're distilled facts, not copies of your conversations.
• Reference chat history. Since April 2025, ChatGPT can also draw on your earlier conversations as a whole when composing an answer. This is best-effort retrieval of what seems relevant — not a transcript replay. It may surface the thing you need; it may not.

So "ChatGPT forgot our previous conversation" almost always means one of two things: the memory layer never captured the detail, or it captured something but didn't judge it relevant enough to bring into the new chat. Neither mechanism promises the fidelity people expect — and that expectation gap is the real answer to the question in the search box.

Why it forgot your conversation: check these in order

Before blaming the model, run down this list. Most cases of "ChatGPT keeps forgetting" turn out to be one of these:

• Memory is switched off. It's on by default for most accounts now, but it can be turned off — check Settings → Personalization → Memory (the older system labels the switch Enable memory).
• You're on a different account. Memory is per-account. A work login on the laptop and a personal login on the phone are two strangers to each other. This one catches more people than any setting.
• It was a Temporary Chat. Temporary Chats don't use your memories, don't create new ones, and don't enter your history. If the earlier conversation was temporary, it left nothing behind — by design.
• You've hit a bound OpenAI doesn't publish in absolute terms. Its Memory FAQ (read 28 August 2026) documents no entry limit and no "memory is full" state for the current memory summary; its release notes do name one for the legacy saved-memories list, in the entry dated 15 October 2025. A bound exists on the summary too, and OpenAI has sized only the change: the release note dated 4 June 2026 says Plus and Pro accounts get "twice as much memory capacity" — twice their previous capacity, whose size OpenAI has never published. If new details have stopped sticking and nothing else on this list fits, this is the likeliest remaining explanation. Here's what OpenAI does and doesn't publish about memory capacity.
• The feature hasn't reached your plan or region. Memory features have rolled out to plans and regions at different times, and the newer summary system is still spreading. Look at what your own Settings screen offers rather than assuming from someone else's screenshot.
• The conversation was deleted. A deleted chat stops being a source ChatGPT can draw on. (Anything it had already distilled into memory persists until you remove it there too.)
• It was never saved in the first place. ChatGPT saves automatically only what it judges useful later. The CFO's exact words about pricing rarely make that cut. If it wasn't told to remember, it probably didn't.

One disambiguation before you go further: if ChatGPT is forgetting things inside a single long conversation — dropping instructions you gave two hours ago — that's a different mechanism entirely (the context window, not memory). We cover that in why ChatGPT forgets things in the same chat.

Even with everything on, it recalls — it doesn't reread

This is the part worth internalizing, because it sets the right expectations for real work. With memory on, the right account, and space available, ChatGPT still doesn't carry your previous conversations forward verbatim. It carries a distillation: a summary of who you are and a best-effort pull from past chats. OpenAI's own help center says the memory summary "will not include everything that ChatGPT remembers" — and there's no complete inventory you can open of what it drew from your wider history for any given answer.

In practice that means memory works well for standing facts — your role, your writing preferences, the names of your clients — and unreliably for specifics: the number the client pushed back on, the deadline agreed three chats ago, the exact wording of a commitment. Those live in the transcript, and the transcript is not what gets carried forward.

How to make ChatGPT hold on to more

Within ChatGPT, five habits meaningfully reduce the forgetting:

• Confirm memory is on — on the account you actually use. Settings → Personalization → Memory, on every device you work from.
• Tell it explicitly. "Remember that I'm the partner on the Acme engagement and the Q3 review is October 14." Explicit requests are the most reliable way into saved memory; automatic capture is a judgment call you don't control.
• Keep long-running work in a Project. A Project keeps instructions and files attached, so chats inside it start from the brief instead of from zero. This is the closest thing ChatGPT has to a durable working file.
• Open new chats with a handoff. Before leaving a thread, ask for a short summary of decisions, constraints, and open items — check it — then paste it at the top of the next chat. Now the important part is the newest thing in the thread, not a memory the layer may or may not surface.
• Review the memory summary occasionally. Skim what it holds, correct what's wrong, delete what's stale. A memory layer you've pruned surfaces better context than one silting up with old projects.

These work. They're also work — a discipline you maintain, per account, forever. Keep that in mind for the last section.

If your saved memories vanished overnight

Searches for "ChatGPT memory wiped" spike whenever people open the memory settings and find them empty. Before assuming the worst, check the mundane explanations: you're signed into a different account than you think; someone on a shared machine used Delete and turn off memory (it clears the summary and is not undoable); or your account moved between the older saved-memories list and the newer memory-summary view — the settings screen shows different things on each, which can read as a wipe. An empty memory summary also has documented benign causes: OpenAI says it can sit empty if your account is new, you only just turned memory on, or there's too little chat history to draw from — it fills in as you use ChatGPT, and the summary's ⋯ menu has a Refresh option. Rule those out before concluding your memories were deleted.

If they're genuinely gone, there's no restore — OpenAI can't recover deleted memories. Two consolations: OpenAI notes that with memory re-enabled, ChatGPT may rebuild some of it from chats still in your history, and you can accelerate that by explicitly re-telling it the facts that matter. But if a morning like that made your stomach drop, it's worth asking why irreplaceable working context was living in a list you can't fully see and didn't back up.

The boundary no setting changes

Suppose you do everything above perfectly. Memory on, one account, Projects for every engagement, handoff summaries, a pruned memory list. You've now optimized recall over one source: your ChatGPT conversations. That's the whole territory. Unless you've handed it your inbox or drive account-wide, it has never seen the thread where the CFO actually pushed back on pricing, the PDF the client sent Tuesday, the voice memo you dictated after the call, or the meeting notes from October — because none of that happened inside ChatGPT.

For client-facing work, that's the forgetting that costs you. The conversation you need before Friday's 9 a.m. isn't a chat you had with a model — it's an email thread, a document, and a note, scattered across three apps. No memory setting reaches them, and pasting everything in by hand just makes you the courier for a system that keeps what it judges useful and shows you part of the result.

An assistant that starts from what you've shared — all of it

Equerry is a personal assistant built on the other side of that boundary. A new ChatGPT chat starts from an empty thread and a selective summary; Equerry answers from everything you've deliberately shared with it, whatever the source:

• CC or forward the threads that matter to your own Equerry address. Once you've CC'd it on a client thread, every later reply lands too — the previous conversation that matters is the one with the client, and it's kept without you re-forwarding each new message.
• Save the rest the way you already capture. Share a PDF or screenshot from the share sheet, save a note, dictate a voice memo after the call. It all lands in one place.
• Ask later and check the source. "What did the CFO say about pricing in the October review?" comes back with the answer and a source chip — tap it to open the exact email or note it came from. No wondering whether a summary caught it.
• Walk in briefed. A morning brief before Friday's call reflects Thursday afternoon's reply you hadn't opened yet — recall that arrives before you ask, not after you prompt.
• See everything it kept — and delete any of it. Every remembered fact is listed, inspectable, and removable. Nothing selective, nothing partial, no saved-memory cap to manage.

ChatGPT's forgetting is fixable at the margins with the habits above, and they're worth adopting. But if what you actually need is an assistant that never opens the conversation asking who the client is, the fix isn't a better prompt — it's keeping your working context somewhere built to hand it back. Start with what ChatGPT does and doesn't keep, or see how Equerry handles the context your work actually runs on.

ChatGPT behavior and control labels above were checked against OpenAI's help center in August 2026, and the memory-capacity statements re-verified on 15 September 2026 against the ChatGPT — Release Notes help-center article (article 6825453), for the entries dated 15 October 2025 and 4 June 2026. help.openai.com refuses automated fetches with a 403, so the release notes were read through a text-extraction reader proxy and confirmed separately in a browser session, both on 15 September 2026. Quoted wording is OpenAI's, and OpenAI publishes no absolute capacity figure on any plan. OpenAI renames and moves these controls often — treat the labels as a guide and check your own Settings screen.

Last updated 2026-09-15.

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.