“The feature most users value is also the one they cannot fully audit or constrain. The memory that makes ChatGPT feel like it knows you is the same mechanism building a behavioural dossier that you are not permitted to see in full.”
ACM CHI 2026 — ‘Relational Gains, Privacy Strains’ · Referenced by Ground TruthIf you are a ChatGPT Plus or Pro subscriber in the United States, something changed on June 4th 2026 that you may not have been explicitly notified about. OpenAI rolled out an upgrade called Dreaming V3 — the most significant overhaul of the platform’s memory architecture since ChatGPT launched. The old system was opt-in and manual: you told it what to remember, it remembered that specific thing. The new system works differently. It reads across your entire conversation history — everything you have ever said to ChatGPT, going back years — and synthesises a persistent profile of you automatically, continuously, in the background, without you prompting it to do anything. Freshness, continuity, relevance: those are the three dimensions the system now tracks. It decides what matters. It decides what it keeps. You can review a summary of what it has stored. The summary, OpenAI acknowledges, may not include everything ChatGPT remembers.
That last sentence is the one that should stop you mid-paragraph. A system is building a profile of you and the company running it has told you, in its own documentation, that the record of what that profile contains is incomplete. This is not a minor technical caveat. It is a significant statement about the limits of the user’s ability to understand, audit, and control what is being stored about them. And it is sitting, largely unremarked upon, in a product update that most people read as “ChatGPT got better at remembering stuff.” Welcome to Episode 02 of Ground Truth. Let us read what everyone else is skimming.
Not a Feature Update. An Architectural Shift.
The name Dreaming V3 is evocative, which is presumably intentional. It suggests something passive, ambient, almost poetic — the AI quietly dreaming about your conversations, making connections, learning who you are. That framing is accurate in a narrow technical sense and deeply misleading in every other sense. What is actually happening is this: the system is performing continuous background synthesis across your full conversation history, extracting what it determines to be signal about your preferences, goals, working style, health mentions, relationships, and behavioural patterns, and storing that extracted signal as a persistent profile that shapes every future response you receive.
For everyday use this is, genuinely, useful. If you have spent months building a project with ChatGPT’s help, you will no longer need to re-explain the context every session. If you have a consistent writing style you want maintained, the system will remember it. If you mentioned six months ago that you are vegetarian, it will not suggest a steak recipe. These are real quality-of-life improvements and they are why the feature has been positively received by most users who tried it.
The problem is not the utility. The problem is the architecture required to deliver that utility. To synthesise context across years of conversations, the system needs to read all of those conversations. It needs to make judgments about what in those conversations is significant enough to retain. It needs to store those judgments in a form that persists indefinitely. And it needs to apply those stored judgments to shape responses without the user necessarily knowing in any given moment which stored memories are influencing what they are being told. That is not a chatbot with a helpful memory. That is a behavioural profiling system with a chat interface.
96% of ChatGPT memory entries in a sample of 2,050 entries from 80 users were created unilaterally by the system — without the user asking it to remember anything. Source: February 2026 arXiv study. ChatGPT Plus and Pro users get twice the previous memory capacity under the new system. The Memory Summary page, OpenAI acknowledges, may not include everything ChatGPT remembers. Free-tier users are next in the rollout queue. 900 million weekly active users. Not all of them read product update notices.
The Profile You Are Building Without Knowing It
A 2026 survey found that 82% of US ChatGPT users considered their conversations sensitive or highly sensitive. That is a striking number given the casual way most people interact with the platform — asking for help with emails, discussing health concerns, talking through relationship problems, working on confidential business documents, processing difficult personal decisions. ChatGPT has become, for many of its 900 million weekly users, something closer to a journal than a productivity tool. People tell it things they would not tell a colleague. They tell it things they would not tell a friend. They tell it things they have not told anyone.
All of that is now being synthesised into a persistent profile that the system retains indefinitely, updates automatically, and applies to shape what you are told in every future interaction. The profile is not just “this user likes bullet points.” It includes inferred goals, inferred anxieties, inferred relationships, inferred health status, inferred political leanings — anything the system determines is signal worth retaining. The ACM CHI 2026 paper “Relational Gains, Privacy Strains” describes this as the personalisation-convenience paradox: the feature users value most is also the one they cannot fully audit or constrain. You get the useful assistant. You also get the dossier. They are the same thing.
There is also a class action lawsuit filed against OpenAI in May 2026 separately alleging that ChatGPT embeds Meta’s Facebook Pixel and Google Analytics tracking code on ChatGPT.com — potentially exposing user queries to advertising networks in real time. OpenAI has not publicly addressed these allegations in detail. The Italian data protection authority fined OpenAI €15 million in December 2024 for GDPR violations related to ChatGPT data processing. European regulators have demonstrated both the willingness and the legal tools to act. The United States has no federal AI privacy law governing consumer chatbot memory as of June 2026.
People use ChatGPT the way they use a diary. They tell it things they would not tell anyone else. All of that is now being synthesised into a persistent profile. You can view a summary of it. The summary may not include everything the system knows. Read that sentence again.Neal Lloyd · Ground Truth, Episode 02
A Memory That Can Be Hijacked Across Sessions
In November 2025, Tenable Research published a demonstration of something that deserves considerably more attention than it received. The research showed that because memories are appended to the system prompt in ChatGPT’s architecture, a maliciously crafted prompt injected via a third-party source — a document you asked ChatGPT to read, a website it was asked to summarise, an email it was asked to analyse — could potentially update persistent memory with attacker-controlled content. The attack creates what Tenable described as an exfiltration channel that survives across sessions.
To translate: someone else’s content, processed by ChatGPT on your behalf, could potentially write false or manipulative information into your persistent memory profile. That poisoned memory then shapes everything ChatGPT tells you in future sessions — not just in the session where the attack occurred, but permanently, until you manually audit and clean your memory store. An attack that persists across sessions, survives conversation resets, and influences the AI’s responses to you without your knowledge is not a theoretical edge case. It is a documented vulnerability in the architecture that Dreaming V3 has just made significantly more consequential, because the memory system is now more powerful, more automatic, and harder to audit than it was before.
OpenAI has not publicly confirmed whether Dreaming V3 addresses this attack surface. On the same day the Dreaming V3 rollout was announced, OpenAI separately introduced Lockdown Mode — an optional security setting that limits web and external service access to reduce prompt injection risk. Lockdown Mode is opt-in. Most users will not enable it. Most users do not know it exists.
You ask ChatGPT to summarise a document. The document contains a hidden instruction. That instruction tells ChatGPT to update your persistent memory with false or manipulative information. ChatGPT complies — because from its perspective, it is just processing a user instruction. That false memory now sits in your profile. It shapes your future responses. You do not know it is there. The Memory Summary page may not show it. This is a documented attack, demonstrated by Tenable Research, published November 2025. Dreaming V3 makes the persistent memory store more powerful and more automatic. The attack surface did not shrink. It grew.
The EU AI Act Privacy Clock Is Ticking. The US Has Nothing.
On August 2nd 2026 — 53 days from the date of Dreaming V3’s rollout — the EU AI Act’s transparency obligations for chatbot systems take effect. Under those obligations, AI systems that build persistent behavioural profiles of users are classified as profiling activities, triggering consent requirements, disclosure obligations, and the right to erasure. OpenAI will be required to meet these standards across its European user base within weeks of deploying a system that, in its current form, does not fully disclose what it retains.
The Italian fine of €15 million established that European regulators have the appetite for enforcement. The GDPR’s right to erasure — the right to have your data deleted — is already technically applicable to ChatGPT memory in the EU. The August deadline makes the compliance requirement more explicit and the enforcement environment more acute. OpenAI is simultaneously preparing a near-trillion-dollar IPO, managing a US regulatory environment that has become notably more permissive under the current administration, and heading into a European regulatory crunch that could require material changes to its most significant new product feature. That is a complicated position to be in eight weeks before a public market debut.
In the United States, there is no federal AI privacy law governing consumer chatbot memory. Colorado’s AI Act — the most consequential piece of US AI regulation actually taking effect in 2026 — has a June 30th enforcement date for anti-discrimination provisions but does not specifically address memory profiling. The Great American Artificial Intelligence Act, the 269-page federal bill dropped by Congress on June 5th, includes a three-year preemption of state AI laws but has not moved out of committee. Companies that delayed compliance planning waiting for federal preemption are, in the meantime, 22 days from Colorado’s real enforcement date with nothing preempting it yet.
Practical Steps. Right Now. Before the Free Tier Rollout Reaches You.
Audit your memory now. Go to Settings → Personalization → Memory in ChatGPT. Read what is there. You will likely be surprised. The Memory Summary represents a partial view of what the system has synthesised about you — OpenAI’s own documentation acknowledges it may not be complete — but it is the only window you have. Delete anything sensitive, inaccurate, or that you would not want influencing future responses.
Understand what you cannot see. The Memory Summary shows you what the system has explicitly stored. It does not show you the full synthesis process — the implicit weighting of your conversation history that shapes responses without being listed as a discrete memory entry. You are auditing the tip of the iceberg. The iceberg exists.
Enable Lockdown Mode if you process sensitive documents. If you regularly ask ChatGPT to read, summarise, or analyse external content — documents, websites, emails — the prompt injection attack surface is real. Lockdown Mode limits this. It is in Settings. It is opt-in. Enable it.
Revert to legacy saved memories if you want explicit control. OpenAI has preserved the option to revert to the manual saved-memories system. Go to Settings → Memory → Saved memories. This gives you explicit, auditable control over what is retained at the cost of the automatic personalisation. For users handling professionally or personally sensitive material, this trade-off is worth considering seriously.
Watch the EU AI Act August 2nd deadline. If OpenAI makes material changes to Dreaming V3’s transparency and consent architecture to comply with European requirements, those changes will tell you a great deal about what the current architecture is missing. European regulatory compliance documents, when they are published, are frequently more informative about a product’s actual data practices than any amount of press releases. Read them.
OpenAI built the most intimate data collection system in consumer technology history and called it a memory upgrade. The marketing is impeccable. The privacy architecture is a work in progress. The 53-day compliance clock does not care about either of those things.Neal Lloyd · Ground Truth, Episode 02
Ground Truth, Episode 02 · June 2026
Neal Lloyd covers the real-world impact of AI — money, power, geopolitics, and the stories behind the headlines. Ground Truth is his live news and analysis companion to the Inside The Machine theory series.
