ChatGPT Memory used to be a simple list: you said “remember this,” it saved a line, and you could read every line back in Settings. That mental model broke on June 4, 2026, when OpenAI started rolling out the Dreaming memory upgrade (internally “Dreaming V3”). Memory is now two layers, and one of them rewrites itself in the background without asking you. A note that read “the user is going to Singapore in July” quietly becomes “the user went to Singapore in July 2026” after the trip — useful when it’s right, and a silent liability when it’s wrong. This guide explains how the June 2026 system actually works, what you can still control, and the 10-minute monthly habit that keeps stored memory from drifting away from the truth.
TL;DR
- Memory has two layers: Saved Memories (an editable, auditable list) and Reference Chat History (implicit recall from past chats, never shown as a fixed list). Both toggle independently at Settings, Personalization, Memory.
- The June 2026 Dreaming upgrade adds a background process that synthesizes and updates memories continuously, without a “remember this” prompt. The new risk is silent revisions — it can quietly correct a detail nobody reviewed.
- Put stable, structural facts (role, answer style, hard rules) in Custom Instructions, not Memory. Custom Instructions never get rewritten by the model.
- For anything sensitive or one-off, use a Temporary Chat — it reads and writes no memory and is deleted within 30 days.
- Review the Memory Summary Page monthly. It’s a high-level view, so it won’t show everything the model inferred, but it’s the only auditable surface you have.
The three places ChatGPT stores facts about you
These are easy to confuse, and the difference decides where each fact should live.
| Surface | Who controls it | Editable / auditable | Best for |
|---|---|---|---|
| Custom Instructions | You only | Yes — two static fields you write | Role, expertise level, answer style, hard rules |
| Saved Memories | You + the model | Yes — a list at Settings, Personalization, Manage memories | Stable facts you want pinned (“I’m vegetarian”) |
| Reference Chat History | The model | No — inferred at runtime, not a fixed list | Evolving context the model picks up on its own |
Custom Instructions are the only fully predictable surface: two fields you write, never rewritten by the model. Saved Memories are a shared list — you can add and delete entries, but as of the Dreaming rollout the model also writes and revises them. Reference Chat History is the slippery one: there’s no list to read, because ChatGPT infers it from your past conversations at the moment it answers.
What changed with Dreaming (June 2026)
Before June 2026, a saved memory only existed if you (or the model, occasionally) explicitly created it, and it never updated — entries just went stale. Dreaming flips that. A single asynchronous background process now reads across many past conversations, synthesizes a memory state, and keeps revising it as your circumstances change. That state isn’t stored in your chat log; it lives in a separate data layer and gets injected into the system prompt at the start of every new conversation. In practice, a fresh chat already “knows” things you never told it in that chat.
Three things follow from that design, and you should plan around them:
- You no longer fully author your memory. The model captures context that arises naturally, so the list grows whether you curate it or not.
- Memories self-update. OpenAI’s own example is the Singapore note rewriting itself after the trip. When it guesses right, that’s convenient. When it guesses wrong, the correction happened silently.
- The audit trail is partial. The Memory Summary Page is, in OpenAI’s words, a high-level view of a broader, continually updated synthesis — so it may not surface every inference that shaped a reply.
OpenAI reports large internal gains from the upgrade (factual recall jumping from roughly 41% in 2024 to ~83% in 2026, by its own evaluation), but those are vendor-stated and unverified. Treat the accuracy as much improved, not guaranteed.
Plan availability and rollout (as of June 2026)
- Initial Dreaming rollout: ChatGPT Plus and Pro in the US, starting June 4, 2026. Plus and Pro get roughly twice the prior memory capacity.
- Coming weeks: Free and Go tiers, enabled by a ~5x reduction in the compute cost of running memory.
- Defaults: Memory is on by default for Free, Go, Plus, Pro, and Team. It is off by default for Enterprise and Edu, where an admin opts in.
- Training: On Free/Plus/Pro, conversations may be used to improve models unless you opt out at Settings, Data controls. Team excludes memory from training; Enterprise and Edu are excluded by default.
If you’re outside the US or on Free/Go and don’t see Dreaming yet, you’re on the older list-based behavior until the rollout reaches you.
What goes where
The split hasn’t changed, even though the machinery under Saved Memories has. Put each fact in the surface that matches how often it changes and whether you need it to be permanent.
| Fact about you | Put it in |
|---|---|
| Your job, location, expertise level | Custom Instructions |
| Preferred answer style (length, tone, format) | Custom Instructions |
| Hard rules (“never write marketing-speak”) | Custom Instructions |
| A stable fact you want pinned (“I’m vegetarian”) | Saved Memories |
| Current project or this quarter’s focus | Saved Memories (and expect Dreaming to keep it) |
| One-off task context | Neither — put it in the prompt |
| Sensitive personal info | Neither — use a Temporary Chat |
The rule of thumb: if a fact is structural and you never want the model touching it, it belongs in Custom Instructions. If it’s true-for-now and you don’t mind the model updating it, Saved Memories is fine. If it’s sensitive or single-use, keep it out of memory entirely.
A 10-minute monthly memory review
Do this on the first of each month. It’s the single habit that keeps Dreaming from quietly accumulating wrong facts about you.
- Open Settings, then Personalization, then Memory. Confirm which layers are on. Toggle off “Reference chat history” if you want only the auditable list to influence replies.
- Open the Memory Summary Page (or Manage memories). Read every entry. With Dreaming active, expect entries you never explicitly created.
- Delete anything outdated, contradictory, or wrong. Click the trash icon next to the entry. Pay special attention to time-sensitive notes the model may have mis-revised.
- Check for anything sensitive. Memory can surface in responses unpredictably, and on shared or account-level access it isn’t private to “your” chats. Delete it and re-run that topic in a Temporary Chat instead.
- Move structural facts to Custom Instructions. “Always answer in concise bullets” or “I’m a senior backend engineer” belong there, where the model can’t rewrite them.
- To wipe everything, scroll to the bottom of the memory list and use “Clear ChatGPT’s memory.” Custom Instructions are separate and survive a memory clear.
A useful tell: if the model cites a fact about you that’s no longer true, that’s your cue to open the summary page and find the stale entry before it spreads to other replies.
Common mistakes
- Assuming you still author the list. With Dreaming, the model writes and revises entries on its own. Curating once isn’t enough; you have to review.
- Trusting the Memory Summary Page as complete. It’s a high-level view. Some inferences from Reference Chat History never appear as a line item.
- Treating Memory like a notes app. It isn’t searchable, isn’t a reliable store, and the model can paraphrase, merge, or silently revise entries. Use a Project file for anything you need verbatim.
- Putting a hard rule in Memory. Rules you never want touched go in Custom Instructions — the model rewrites Memory, not Custom Instructions.
- Using a Temporary Chat for something you wanted kept. It writes no memory and is deleted within 30 days. Nothing carries over.
- Sharing an account and assuming privacy. Memory is account-level, not per-chat.
FAQ
- What’s the difference between Saved Memories and Reference Chat History? Saved Memories is an editable, auditable list you can read in Settings. Reference Chat History is implicit — the model infers context from past chats at runtime and never shows it as a fixed list. You can toggle each one off independently.
- Can I stop ChatGPT from updating memories on its own? Not the Dreaming synthesis itself, but you can turn Memory off entirely, toggle off Reference Chat History, or delete any entry it creates. For a session that touches nothing, use a Temporary Chat.
- Why did it “remember” something I never told it to save? That’s Dreaming. The background process captures context that came up naturally across chats, without a “remember this” instruction. Delete anything you don’t want kept.
- Can I export my Memory? There’s no official export. You can read the Saved Memories list and copy it manually; Reference Chat History has no list to export.
- Is my Memory used to train models? On Free/Plus/Pro it may be, unless you opt out at Settings, Data controls. Team, Enterprise, and Edu exclude memory from training by default.
- Do Custom GPTs see my personal Memory? Generally no. Custom GPTs run on their own system message and don’t pull your personal memory.