Why does ChatGPT forget what I just told it?

Two possibilities. Inside the same chat: the context window is filling and the client is summarizing older turns to make room. Across chats: there's nothing carrying that fact forward unless it got pinned to memory. If you said "remember X" and it didn't show up next chat, either memory was off or the budget was full.

Because Claude has no persistent cross-chat memory by default. Full stop. If continuity matters, use a Claude Project with a "Working memory" doc you keep updated, or paste a briefing at the start of each session. Details: back up Claude Projects.

The conversation transcripts do carry between sessions — they're saved in your chat history. What doesn't carry is the model's active awareness of them. Even with "Reference chat history" on, the model gets a small summary, not the full text of past chats. Turn the toggle off and it gets nothing.

Almost always one of three: a memory reset from a model update, an account issue (email change, region switch, suspension), or the memory budget hit its cap and evicted older facts in a batch. Check Settings → Personalization → Memory — if the list is short or empty, you've been reset.

Style is inferred from context, not stored. Even with memory on, "writes in short punchy sentences" is a stored fact — but the actual prose style is regenerated fresh each time from whatever's in the current context. The fix is usually a custom instruction plus a pinned example, not more memory. See custom instructions vs memory.

Your name is one of the first things ChatGPT will pin to memory if you say it explicitly. If it keeps forgetting, either memory is off, the pin got evicted, or you're on a temporary chat. Say "Remember that my name is …" once with memory on, then check the memory list to confirm.

Project context sits in the current chat's window, not in long-term memory. When the chat closes, that context is gone — even if the chat itself is saved. Move recurring project context into a Custom GPT (system prompt), a Project (attached files), or pinned memory facts, not into the conversation itself.

Different failures. Forgetting: the model doesn't reference a fact it used to. Hallucinating: the model states a fact it never had. If ChatGPT confidently misremembers something, it's hallucinating — filling a gap with plausible text. See hallucinations about you.

Two numbers. In one conversation: 128K tokens on GPT-5 Instant, up to 400K on GPT-5 Thinking (roughly 300–800 pages of text). Across conversations: 1,200–2,000 tokens of pinned facts on the memory feature — a few paragraphs total. Deep dive: ChatGPT memory limits.

200K tokens on Sonnet 4 and Opus 4 — about 500 pages of text. That's per conversation, not per account. When you hit it, Claude silently drops the oldest turns. There is no cross-chat memory to compete for the space.

The pinned-memory budget is a fixed size. When you (or the model) try to add a new fact and there's no room, ChatGPT surfaces the "memory is full" nudge and asks you to delete something. Free tier fills faster than Plus. Full walkthrough: ChatGPT memory full.

Pinned facts persist until you delete them, memory hits its cap and evicts them, or a model update resets the store. None of those are "forever." If a fact matters, back it up outside ChatGPT.

No. A bigger window makes the present larger — it doesn't make the past retrievable. Even with a 1M-token window, the model still sees zero across sessions unless you re-inject the past. Fix has to live outside the model.

A token is roughly 4 characters of English text — "the" is one token, "extraordinary" is three. Models count in tokens because that's what they actually process. A "message" can be one token or ten thousand, so message counts wouldn't mean anything.

Roughly 200K tokens across all attached files plus the current conversation. If you attach a 150K-token PDF, you only have 50K left for actual talking. Trim aggressively; most Projects have too much material for the model to actually use.

Three usual causes. (1) The memory feature inferred it from other chats and pinned the inference as fact. (2) A summarized older turn compressed away nuance. (3) A hallucination. Check Settings → Personalization → Memory and delete anything wrong — those pins actively influence future replies.

Because it has no shared memory to check against. Ask twice in different chats and you'll often get two answers. Even in one chat, contradictions creep in as older statements get summarized. See AI memory contradictions.

Because the model always produces a fluent answer, even when it has no data. If you ask "what was that thing I mentioned last week?" and last week isn't in memory, it will confidently invent something. The fix is to give it a real source (paste the note, or use a memory tool that retrieves) rather than trust its recall.

Two forces. Model updates change the default voice underneath your instructions. And the model regenerates personality from context each turn — so as older context drops out, the voice drifts. See persona drift.

Usually one of: temperature was cranked (more randomness), memory pulled a wrong inference, the current context is confusing the model, or a safety filter half-fired and garbled the response. See why AI says weird things.

Because it stores flat facts, not a graph. It might have "user has a sister named Jane" and "user has a friend named Jane" and quietly conflate them. Long-form: why AI misremembers relationships.

Memory creep. The model infers facts from context — location from your projects, age bracket from your tone — and sometimes pins those inferences. Not malicious, but worth auditing. See AI memory creep.

Because pinned memory doesn't age out on its own. A job you left two years ago stays pinned until you tell ChatGPT to update it. Open the memory list, delete stale pins, and state the current version explicitly.

Turn on Settings → Personalization → Memory (both toggles). Then either wait for it to save facts on its own, or say "remember that…" explicitly. Only the saved-memories list is human-editable — the chat-history summary is opaque.

Settings → Personalization → Memory → Manage. You'll see the list of pinned facts. You can delete individual items or clear the whole thing. The chat-history-derived summary isn't shown, which is one of ChatGPT memory's weaker parts.

Two steps. (1) Clear pinned memories via Settings → Personalization → Memory. (2) Turn off "Reference chat history" and delete past chats (or archive them). For complete removal, do a data export first, then request account deletion. Full guide: delete what AI knows about you.

Yes. Settings → Data Controls → Export data. OpenAI emails a ZIP with every conversation and a memory.json of pinned facts. Link expires in 24 hours. Walkthrough: how to export ChatGPT memory.

Yes. Settings → Privacy → Request data export. You get JSON of every conversation. There's no separate memory export because there's no persistent memory to export. See back up Claude Projects.

Not natively — the two systems don't talk. But you can export from ChatGPT, distill the important atoms, and paste them into a Claude Project as a working-memory doc. Step by step: transfer ChatGPT memory to Claude.

Export monthly from every provider you use, store the ZIPs outside the provider (Dropbox, iCloud, local drive), and — if you actually want to reuse the memory — load them into a tool that can distill and re-inject. Full playbook: complete AI backup guide.