privacy

ChatGPT's Dreaming Memory: What It Stores About You

ChatGPT's Dreaming memory system builds a persistent profile from every conversation. Here's what it retains, what you can't see, and how to take back control.

OpenAI launched a feature called Dreaming in June 2026, and it quietly changed how ChatGPT handles everything you tell it. Dreaming is not a list of facts you manually save. It is a persistent, self-updating profile that learns from your conversations, stores inferences, and carries them forward — even after you delete the chat.

If you use ChatGPT for anything personal — journaling, health questions, work brainstorming, family planning — Dreaming may know more about you than you realise.

What Dreaming Actually Does

Before Dreaming, ChatGPT’s memory was explicit. You told it to remember something, or it asked, and a fact was saved to a visible list. The list was imperfect, but it was readable. You could look at it and know what the system held.

Dreaming replaced that model with a system that runs continuously in the background. As you chat, it identifies details worth retaining — your job title, your family situation, your health concerns, your anxieties — and stores them without you doing anything. It builds what OpenAI calls a “living profile.”

That profile persists across all your conversations. The ChatGPT you open tomorrow already carries forward what you discussed last month.

The feature launched to paid subscribers in mid-2026 and has since extended to free accounts in certain regions.

The Privacy Concerns Worth Understanding

1. You Cannot See Everything It Stores

OpenAI has acknowledged that the Memory Summary Page — where you can review saved memories — “may not contain all memories.” The full profile is not fully visible to users.

This is not a minor footnote. If you cannot see what is stored, you cannot correct errors, you cannot make an informed judgment about what you want retained, and you cannot act on your right to deletion with any confidence that the deletion is complete.

2. Deleting a Conversation Does Not Delete the Memory

This is the detail that catches most users off guard.

When you delete a chat, the chat history is removed from the History panel. The memories Dreaming extracted from that conversation are not removed. They live in a separate data store with a separate deletion flow.

To remove memories, you have to navigate specifically to Settings → Personalisation → Memory → Manage memories, and delete entries there. The two flows — chat history and memory profile — do not sync. One action does not trigger the other.

3. Dreaming Infers — It Does Not Just Record

A traditional saved-fact memory system stores explicit statements: “I live in Berlin,” “I have two kids,” “I’m vegetarian.” Dreaming’s design goes further. It reportedly draws inferences from patterns across your conversations.

Ask enough questions about managing anxiety, and it may conclude you experience anxiety — and store that as a profile trait. Ask about managing a family member’s dementia, and it may infer details about your personal life well beyond what you stated.

Inferred data about your mental or physical health, your financial stress, your political inclinations, or your family structure sits in a profile you did not consciously create.

4. The Feature Is On by Default

Dreaming is opt-out, not opt-in. Many users who updated to newer ChatGPT builds in mid-2026 inherited the feature without actively choosing it. The toggle to disable it requires navigating Settings → Personalisation → Memory — not a prominent option in the main interface, and not something the product draws attention to.

The pattern of defaulting new privacy-touching features to “on” is worth noting. It shifts the burden of protection to users who may not know there is anything to protect.

What the Data Could Be Used For

For now, OpenAI describes the Dreaming profile as serving personalisation — making responses more relevant by carrying context across sessions. The data is retained on OpenAI’s servers subject to their standard data retention and deletion policies.

That creates several categories of downstream exposure:

Law enforcement requests. Like any cloud service holding user data, OpenAI is subject to valid legal requests for user data. A Dreaming profile — containing months of inferred health details, financial stress signals, relationship concerns — is a significantly richer subject of a legal request than a simple list of saved facts.

Data breach. The richer the persistent data store, the more valuable the contents if those servers are ever breached.

Policy changes. OpenAI’s data retention and training policies have changed before. What is stored for personalisation today could serve different purposes under a future policy revision. The data exists because it was retained; its uses are governed by terms of service documents that can be updated.

Third-party AI access. Users who connect OpenAI services to other tools — through plugins, GPT Actions, or API integrations — may expose their profile data to those integrations. The Dreaming profile is context that travels with those connections.

What You Can Do Right Now

These steps take about ten minutes and should be done once, then revisited periodically.

Step 1: Review your current memory profile. Go to Settings → Personalisation → Memory → Manage memories. Read every entry. Delete anything you would not want stored on a stranger’s server.

Step 2: Delete the full profile. There is a “Clear all memories” option at the bottom of the Memory management page. Use it to start fresh.

Step 3: Turn Dreaming off. The same Settings → Personalisation panel has a toggle to disable memory entirely. This prevents new entries from forming. Do this after clearing the existing profile, not before.

Step 4: Use Temporary Chat for sensitive topics. ChatGPT’s Temporary Chat mode (accessible from the sidebar) starts each conversation fresh — no memory access, no contribution to the Dreaming profile. Use it for any conversation you would not want archived.

Step 5: Set a calendar reminder. Re-examine your memory profile every 60–90 days. If you re-enable the feature or forget you disabled it, the profile can silently rebuild.

The Broader Pattern: Persistent AI Memory Is Becoming the Default

Dreaming is the most visible instance of a wider shift. Persistent memory is becoming a standard feature, not an experiment.

Google’s Gemini rolled out “Personal Context” as a default in 2025, drawing from your Gmail and Google Drive to inform responses. Siri in iOS 27 carries contextual data across Apple apps and — in some configurations — to third-party AI providers selected by the user. Microsoft Copilot logs interactions across your Microsoft 365 account.

Each of these is individually manageable, if you know where the settings live. Together, they mean your questions, worries, plans, and personal details are being archived across a growing list of platforms. Each platform has its own retention policy, its own deletion flow, and its own legal exposure.

The AI assistant tools most people use daily are, by design, building detailed models of who you are.

A Different Model: Storage That Doesn’t Read What You Store

One practical response is to separate the tools you use for AI interaction from the tools you use for personal storage.

AI assistants get better by knowing more about you. Personal file storage does not need to know anything about you beyond how to authenticate and serve files back.

daftei stores photos, voice notes, documents, and memories with AES-256 encryption at rest and TLS 1.3 in transit. It is GDPR and CCPA compliant. It never uses your stored content to train AI models — third-party or otherwise — and it never shows ads. Your files are there to be retrieved by you, not analysed.

The two categories of tools can coexist. The important thing is to know which category you’re in when you type something personal.

Three Questions to Ask Any AI Memory Product

Before you accept the default settings on any AI memory feature, three questions are worth looking up in the product’s documentation:

1. What exactly gets stored? Is it verbatim statements, inferences drawn from patterns, or both?

2. Does deleting a conversation delete the associated memories? If not — and in Dreaming’s case the answer is no — what is the separate deletion flow and how complete is it?

3. Who can access the stored profile? Can employees access it for quality review? Is it accessible on a valid law enforcement request? Does it travel with third-party integrations?

If the documentation does not clearly answer all three, that is informative in its own right.

The Trade-off Is Real

AI memory features are genuinely useful. Context across sessions saves time, prevents repetition, and makes responses sharper. The trade-off is not between convenience and paranoia — it is between personalisation and control over a profile you may not be able to fully see or delete.

Dreaming represents one of the largest expansions of passive data collection in consumer AI so far. Understanding it is not optional if you use the product for anything personal.

The controls exist. They are just not where most users think to look.

Your memories deserve better than an ad platform.

Try daftei free →
← All posts