ChatGPT's memory now rewrites itself. Here's what that costs you
Until June 2026, you could open ChatGPT's settings and read a list of what it had saved about you: discrete entries, each one added when you said "remember this" or when the model decided something was worth keeping. You could delete any line. That list is gone. OpenAI's Dreaming update, announced June 4, 2026, replaced it with a memory that continuously rewrites itself in the background, and a summary page that shows you some of the result.
The recall is genuinely better. The audit trail is genuinely worse. I went through the current vendor documentation for all four major assistants to see which of them can still answer a simple question: what did this thing believe about me last month?
Quick answer: Since the June 4, 2026 Dreaming update, ChatGPT's memory is a continuously self-rewriting synthesis of your chat history, not a list you manage. You can correct or suppress items on the memory summary page, but there is no revision history, and OpenAI says the page does not show everything. A complete audit is no longer possible.
What changed with ChatGPT's memory in June 2026?
Dreaming replaces the saved-memories list as the primary storage with a background process that reads across your past conversations and updates what ChatGPT remembers without being asked. Instead of waiting for "remember this," it picks up context as it appears. OpenAI surfaces the result on a memory summary page, where you can read highlights, add or correct information about yourself, and give instructions about what to bring up and when. The update reached US Plus and Pro subscribers first, with Free and international users following.
Three things disappeared in the trade, as covered in detail by XDA's June 2026 analysis and TechTimes:
- The complete list. OpenAI's help documentation acknowledges the memory summary does not include everything ChatGPT remembers, and that the system itself decides which details are appropriate to show. The page built for reviewing your memory is a partial view by design.
- Durable per-item deletion. You can delete a specific saved memory (separately from the "Don't mention this again" suppress option), but because the summary continuously re-synthesizes from your chat history, a deleted line can be re-derived. OpenAI's own guidance is that fully removing something means deleting the memory entry and the conversations it came from.
- Any notion of history. When the background process revises what it believes about you, the old belief is simply gone. There is no record of what changed, when, or why.
Why does a self-rewriting memory matter?
For most everyday use, it doesn't. A memory that maintains itself is more useful than one you have to garden, and OpenAI's bet is that most people would rather have better recall than a management interface. That bet is probably right.
It matters the moment you need to answer a question about the memory itself. Three examples:
- "What did it believe about me in March?" Unanswerable. The summary shows current state only. If Dreaming quietly corrected a wrong detail, or introduced one, nobody reviewed that change and no trace of it exists.
- "Is this detail actually gone?" Hard to say. Suppression hides a memory from view; the reporting on Dreaming and OpenAI's own partial-view caveat make it difficult to confirm what is retained underneath.
- "Can I hand this to my DPO?" Not in a form they can rely on. You can manually copy the memory list out of Settings, but it is unstructured text, not an auditable record. ChatGPT's account-level data export (Settings → Data Controls) covers your chats, not a structured memory state, as of July 2026.
If you work under confidentiality duties or GDPR obligations, note the shape of the problem: rights of access and rectification assume there is a stable record to access and rectify. A synthesis that rewrites itself between your review and your correction is a moving target. That is not automatically a violation of anything. It just makes "demonstrate what the system knew and when" impossible with the tools provided.
How do Claude, Gemini, and Microsoft Copilot handle memory?
All state below is as of July 2026, checked against each vendor's own documentation.
Claude
Claude changed its memory on July 10, 2026, and the new version is closer to what this article argues for. It now keeps "a set of individual entries that are organized into categories", updated in real time rather than synthesized once a day, and you can open any entry and delete it (Free, Pro, and Max are on the new model; Team and Enterprise are still migrating). Anthropic also ships an experimental memory import/export for moving memory between Claude and other AI services. What it still does not have: provenance for any given entry, or history of how one changed. Anthropic's own documentation is explicit that "Individual user memory edits are not logged." The entries are also decoupled from their sources: deleting a conversation does not remove memories generated from it, which is the exact inverse of Dreaming's re-derivation problem, and worth understanding on its own. For a walkthrough, see how to audit what Claude remembers about you.
Gemini
Gemini keeps a saved-info page listing things you have told it plus details it has picked up, and you can view, edit, or delete entries there, alongside broader personalization from your chat history. Closest of the big three to a browsable list. Still no change history.
Microsoft 365 Copilot
Copilot's memory (in preview) stores saved memories, inferred details, and custom instructions in a hidden folder in your Exchange mailbox. You can delete saved memories under Settings → Personalization. But Microsoft's own admin documentation is unusually frank about the governance gaps: details inferred from chat history are "dynamic" and Copilot "might update or discard older details" on its own; Purview retention policies do not apply to Copilot memory; and, verbatim, "Memory and personalization actions don't generate audit log entries in Purview." An enterprise compliance suite whose AI memory is exempt from its own audit logging is worth reading twice; I dug into what all four vendors actually log in your AI has an audit log now.
note: The pattern across all four vendors: view and edit exist everywhere, in some form. Revision history, per-fact provenance, and validity over time exist nowhere. ChatGPT's Dreaming update moved the market leader further from auditability, not closer.
What should an auditable AI memory actually look like?
Auditable memory is a solved problem in other domains. Accounting systems, medical records, and version control all settled on the same principles decades ago. Applied to AI memory, five properties:
1. Individual facts, not a synthesized blob
A record you can point at: "works at Acme," "prefers morning focus blocks." You cannot audit a paragraph that summarizes an unknown number of underlying beliefs.
2. Provenance on every fact
Which conversation, note, or event produced it. When a memory is wrong, provenance tells you whether the source was wrong or the extraction was.
3. Validity ranges, not overwrites
When you change jobs, "works at Acme" should not be erased. It should be closed (true from 2023 to 2026) and a new fact opened. This is bi-temporal storage: every fact carries when it was true and when the system recorded it.
4. Revision history
Every change is kept. The question from earlier, what the system believed about you in March, becomes a query you can run.
5. A consent gate and an exit
New memories held in a pending-review queue until you approve them, and the whole store exportable in a structured format at any time.
Calmara is one implementation of exactly this: memory stored as plain-English subject/predicate/object facts with bi-temporal validity ranges, a pending-review queue so nothing is remembered without approval, and full export on every tier, including free. The six controls your compliance team needs walks through why each property matters if you answer to a regulator. The honest tradeoff: this is heavier than a self-maintaining summary. If you never need to prove anything about your AI's memory, Dreaming-style synthesis is more convenient, and you should use it with clear eyes about what it cannot tell you.
What can you do about ChatGPT's memory today?
- Read the memory summary page. Settings → Personalization → Memory. Correct anything wrong now; you cannot see what it said last month. (A versioned memory can answer that — what a time-travel query looks like.)
- Know the difference between suppress and delete. "Don't mention this again" suppresses a detail; the ⋯ menu deletes the entry. Neither is durable on its own: as long as the source conversations exist, re-synthesis can bring the detail back. If something must not exist in the memory at all, delete the memory entry and the chats it came from, or keep memory off.
- Use temporary chats for anything that should never be learned. Client details, health questions, anything confidential. Temporary chats are excluded from memory.
- Export your account data periodically. Settings → Data Controls → Export. It captures chats rather than memory state, but it is the only paper trail on offer.
- Keep the authoritative record elsewhere. If what an AI knows about your work has professional consequences, the memory that matters should live in a system built for audit, and your assistant should read from it. That is the architecture behind an AI second brain you can actually audit.
FAQ
Did the June 2026 update delete my old saved memories?
No. Existing memories were folded into the new synthesized state. What disappeared is the list view of them as individually managed entries; the summary page shows highlights of the combined result.
Can I export ChatGPT's memory?
There is no structured memory export, but you can copy it out by hand: open Settings → Personalization → Memory → Manage memory in a web browser and select-all-copy the list. The account-level data export (Settings → Data Controls) includes your conversation history but not the memory state. A full walkthrough, including where to put the copy so it stays auditable, is in how to migrate your ChatGPT memory.
Can I still turn ChatGPT's memory off entirely?
Yes. Memory can be disabled in Settings → Personalization, and temporary chats bypass it per conversation. Turning it off is the only complete opt-out from background synthesis.
Does Claude have the same problem?
Partly. Claude also keeps no revision history, but since July 2026 it stores individual entries you can open and delete, and it can be paused, reset, or moved with an experimental import/export. Details in the Claude memory audit guide.
What does bi-temporal memory mean?
Every fact carries two timelines: when it was true in the world, and when the system recorded it. Corrections close the old fact instead of overwriting it, so the full history of what was believed, and when, stays queryable.
Is a self-rewriting memory a GDPR problem?
Not automatically. GDPR grants you access to your personal data and the right to correct it, and the summary page arguably serves both. The friction is demonstrability: a record that rewrites itself without history is hard to verify, for you or for the controller. If you process client data under GDPR yourself, that difficulty becomes your difficulty.
What is the difference between suppressing and deleting a memory?
Suppression ("Don't mention this again") stops ChatGPT from surfacing a detail; deletion removes that saved-memory entry. ChatGPT offers both, but neither is durable while the source conversations remain, since the summary re-synthesizes from your chat history. The full breakdown of what "remove" actually does in each assistant is in suppressed or deleted?.
Why did OpenAI make this change?
Better recall. OpenAI's stated goal is memory that captures context naturally instead of waiting for explicit "remember this" instructions, and coverage of the update reports meaningfully improved recall quality. The design trades transparency for capability. Whether that trade is right depends on what you use it for.
---
Read your assistant's memory page this week, whichever assistant it is. If what you find there matters professionally, the follow-up question is whether you could prove it a month from now. For the version of this problem where the answer must be yes, start with how to audit what Claude remembers about you.
Written by Dan Hagen