OpenAI began rolling out Dreaming V3 on June 4, 2026, a new ChatGPT memory architecture that replaces the manually curated saved-memories list with a background synthesis process that reads across years of past conversations and updates what the system remembers about a user without any prompting. The update reached Plus and Pro subscribers in the United States first; Free, Go, and international users are expected to follow in the coming weeks. For hundreds of millions of ChatGPT users, this means the assistant is about to know more about them than most realize — and the controls for managing that knowledge are not as simple as a single toggle.

The core change is architectural. Until June 4, ChatGPT memory relied on two distinct layers: an explicit list of saved facts a user had either stated or confirmed, and Dreaming V0 — a background process introduced in April 2025 that could reference broader chat history but was never capable of standing alone as a memory system. Dreaming V3 collapses that structure. A single asynchronous background process now synthesizes memory from many conversations simultaneously, automatically captures context that arises naturally in conversation, and updates existing memories as circumstances change. OpenAI's own example: a memory reading "you're going to Singapore in July" rewrites itself to "you went to Singapore in July 2026" after the trip ends, with no action required from the user.

That temporal awareness is the headline capability. It is also why the feature demands more user attention than a static list of saved facts ever did.

How ChatGPT Dreaming V3 Background Synthesis Actually Works

The memory state produced by Dreaming V3 is not stored inside the ChatGPT conversation log. It is maintained in a separate data layer and injected into the system prompt at inference time — meaning every new conversation starts with a context window that already contains what the system has synthesized about the user from past sessions. Tenable Research documented this architecture in November 2025 while investigating seven ChatGPT vulnerabilities: because user memories are appended to the system prompt, a maliciously crafted prompt injected through a third-party source — a document, a linked webpage, a tool output — can instruct ChatGPT to update persistent memory, creating an exfiltration channel that survives across sessions. OpenAI has not disclosed whether Dreaming V3's architecture specifically addresses this attack surface.

The separation of memory storage from conversation logs has a direct practical consequence: deleting a conversation does not remove memories derived from it. Saved memories are stored in a separate system; removing the source conversation leaves the synthesized output intact. To fully remove a detail, a user must delete both the memory entry and the originating conversation. Even then, OpenAI's Memory FAQ states that logs of deleted saved memories may be retained for up to 30 days for safety and debugging purposes.

The roughly 5x compute reduction OpenAI credits for enabling the free-tier rollout reflects an internal architectural optimization in how the background synthesis process is batched and served at scale. OpenAI has not published the specific method — whether from distilled memory models, more efficient inference scheduling, or other changes — but the result is that a process previously too expensive to run for free-tier users is now practical to deploy to the entire ChatGPT user base. For Plus and Pro users, the same efficiency gain translates into 2x more memory capacity.

ChatGPT AI Memory Privacy Controls: What Each Toggle Actually Does

ChatGPT memory has two controls that behave differently, and conflating them is the most common user error.

Turning off saved memories also disables reference to chat history. This is a single switch with two effects. When both are off, Dreaming V3's synthesis process no longer draws from past conversations, and OpenAI says information already synthesized is generally deleted within 30 days — though that deletion is not immediate.

Temporary Chat provides a stronger boundary. Conversations held in Temporary Chat mode do not use or update memory, making them the appropriate tool for one-off sensitive discussions — a medical question, a financial situation, a private project — where cross-context synthesis is unwanted.

A third control operates entirely separately: the "Improve the model for everyone" setting, which governs whether conversations may be used for model training. Turning off memory does not disable this setting. Users who want both — no persistent memory and no training contribution — must adjust each control independently.

OpenAI's Memory Summary Page, introduced alongside Dreaming V3, shows what the system has synthesized about a user and allows corrections, dismissals, and topic restrictions. However, OpenAI acknowledges the summary does not necessarily include everything ChatGPT may remember from past conversations. Selecting "Don't mention this again" within the summary reduces future references to a detail but does not delete it; the underlying entry remains in the memory data layer unless explicitly removed.

OpenAI Memory Update Reaches Free Tier: What Changed in Three Generations

ChatGPT memory has gone through three distinct architectures in two years.

The first, launched in April 2024, was a notepad system. A user told it what to remember, it saved a line item, and that entry stayed frozen until changed. The system was brittle: it retained nothing it had not been explicitly instructed to keep, and everything it saved eventually went stale.

The second, Dreaming V0, launched in April 2025 and added a background layer that could reference chat history beyond the saved list. It improved personalization meaningfully — OpenAI reports factual recall climbing from roughly 41.5% to approximately 67.9% between the two architectures — but the company was candid that V0 was never capable of standing on its own. These figures are vendor-stated and have not been independently verified.

Dreaming V3 removes that dependency. Memory now rests entirely on the background synthesis process. Factual recall in OpenAI's internal evaluation reached 82.8% with the 2026 architecture; preference adherence scored 71.3% and time-sensitive accuracy 75.1%. These figures are OpenAI's own and have not been confirmed by an independent auditor.

The scope of what the system can now infer has expanded accordingly. A 2026 study published at the ACM CHI Conference — "Relational Gains, Privacy Strains: Exploring Users' Perceptions and Experiences with ChatGPT's Memory Feature" — found that persistent synthesis of user profiles creates what the researchers described as a "personalization-convenience paradox": the feature most users value is also the feature most users cannot fully audit or constrain.

Regulatory Pressure Surrounds Dreaming V3 Launch

The Dreaming V3 announcement did not land in a calm regulatory environment. Under the General Data Protection Regulation, AI systems that build persistent behavioral profiles of users are classified as profiling activities, triggering consent obligations and the right to erasure. The EU AI Act's transparency obligations for chatbot systems are scheduled to take effect on August 2, 2026 — less than two months after the Dreaming V3 rollout — meaning OpenAI will need to meet new disclosure and data-governance standards within weeks of deploying its most ambitious memory architecture yet.

A class action filed against OpenAI in May 2026 separately alleged 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 without adequate disclosure. The Italian data protection authority fined OpenAI €15 million in December 2024 for GDPR violations related to ChatGPT's data processing, establishing European regulators' readiness to use existing enforcement tools against AI products. The United States has no federal AI privacy law governing consumer chatbot memory as of June 2026.

User-Side Confusion: Cross-Context Personalization vs. Data Leaks

Online discussion since the rollout has included the term "context bleed" to describe cases where information from one conversation surfaces in an unrelated one — for example, a health detail from one session shaping a dietary recommendation in another, or a resolved financial concern continuing to influence responses months later.

That term requires careful handling. Context bleed in this context does not indicate a leak of data between different users' accounts. It describes cross-context personalization within a single account behaving in ways the user did not anticipate. The system may be functioning precisely as designed; the tension is between OpenAI's judgment of relevance and the user's expectation of context boundaries.

That distinction matters because ChatGPT users regularly discuss sensitive subjects. A 2025 survey of 300 US ChatGPT users found that 82% of respondents considered their chatbot conversations sensitive or highly sensitive. A separate analysis of Reddit posts identified recurring patterns of privacy concern, informal boundary-setting, and users working around memory behavior rather than adjusting settings directly.

Dreaming V3 does not mean ChatGPT secretly retains everything, and no evidence of cross-user data leakage has emerged. What it does mean is that personal context — health, finance, relationships, ongoing projects — can now travel further across conversations than many users have understood the system to do. The controls exist. They require deliberate configuration to work as most users would expect.

Frequently Asked Questions

What is ChatGPT Dreaming and how does it work?

ChatGPT Dreaming is a background memory synthesis process that automatically consolidates context from many past conversations and injects it into the system prompt at the start of each new chat. Unlike earlier saved memories, which required a user to explicitly instruct ChatGPT to remember something, Dreaming V3 captures context that arises naturally in conversation and updates existing memories as circumstances change — for example, revising a memory about an upcoming trip to reflect that the trip has now passed — without any user action.

How do I delete ChatGPT memory?

Navigate to Settings > Memory and find the entry in the Memory Summary Page; select it and delete it directly. Also delete the original conversation if you want to eliminate both the source and the synthesized entry. Turning off the reference chat history setting triggers deletion of synthesized memories generally within 30 days — that is not immediate — and logs of deleted saved memories may be retained for up to 30 days afterward for safety and debugging.

Is ChatGPT memory private?

Memory is account-scoped — one user's memories are not accessible to other users. However, OpenAI acknowledges the Memory Summary Page does not show everything the system has retained, and memories stored in the separate data layer persist after conversation deletion. The memory control and the model-training control are separate settings; disabling memory does not automatically opt a user out of conversation-based model training.

What is the OpenAI memory update for free users?

The June 4, 2026 update extends Dreaming V3, OpenAI's background memory synthesis architecture, to Free-tier ChatGPT users for the first time. The expansion was made possible by a roughly 5x reduction in the compute cost of serving the dreaming process at scale. Free users receive a version of Dreaming V3 that meets OpenAI's quality bar; Plus and Pro users receive the additional benefit of 2x expanded memory capacity.