ChatGPT’s memory has always been a snapshot, a record of what you told it, frozen at the moment you told it. That’s the problem the “Dreaming” update is designed to solve.
Available as of June 4, 2026, to ChatGPT Plus and Pro subscribers in the United States, Dreaming is described by OpenAI as a memory synthesis system that curates stored information for “freshness, continuity and relevance.” Third-party reporting from ng.investing.com refers to the update as “Dreaming V3,” though OpenAI’s own materials use “Dreaming”, the version designation hasn’t been confirmed from a primary source.
The core mechanism, according to OpenAI, is temporal revision. The system is designed to update stored memories as time passes. OpenAI illustrates this with an example: a trip reference stored as “going in July” would be revised to “went in July 2026” once the date passes. The specific example originates from OpenAI’s own materials and hasn’t been independently verified, but it’s consistent with what third-party coverage describes, a system built to address the accuracy problems that come from memory that doesn’t age well.
OpenAI describes Dreaming as designed to manage user context over extended, potentially multi-year interaction histories. That’s the claim. Independent evaluation doesn’t exist yet, so treat it as a design goal, not a confirmed capability.
ChatGPT Memory: Before and After Dreaming
Unanswered Questions
- How does the autonomous revision mechanism interact with enterprise data retention policies?
- What audit trail exists when Dreaming revises a stored memory that previously influenced model behavior?
- Does the update affect memory behavior for API-connected workflows, or is it consumer interface only?
The part nobody mentions
is what this means for practitioners building on top of ChatGPT’s consumer workflows. Memory persistence across sessions introduces a data-retention variable that many enterprise deployments haven’t planned for. When the model starts revising what it remembers, autonomously, the audit trail for why it behaves differently in session 500 versus session 5 gets harder to reconstruct. That’s a practical consideration OpenAI’s rollout materials don’t appear to address directly.
For users who want control, help.openai.com confirms the existing memory management options remain in place: you can delete individual memories, clear all memories, or turn memory off entirely. That’s meaningful for privacy-aware users, and it’s the right thing to mention alongside the new capability.
The rollout is US-only at launch. OpenAI has indicated broader availability, including global access and free-tier inclusion, is expected in the coming weeks, though no specific date has been confirmed. Don’t build a deployment timeline around “coming weeks.”
ChatGPT Plus runs $20 per month; Pro runs $200 per month, per current help.openai.com pricing. Note: that pricing structure may include additional tiers, the plan lineup has been evolving, and those figures reflect what’s T1-confirmed as of this writing, not necessarily the full current offering.
What to Watch
This update doesn’t arrive in isolation. OpenAI’s move toward temporal, self-revising memory is part of a broader competitive dynamic where frontier AI labs are treating persistent user context as a differentiation layer, alongside similar features from Gemini and Claude. The question isn’t whether this is useful. It’s whether the autonomy of the revision mechanism, combined with extended interaction histories, creates compliance surface area that enterprise teams haven’t mapped yet.
Wait for independent evaluation of the temporal revision mechanism before treating it as a reliable production behavior. And if your organization uses ChatGPT for any knowledge-work application, now is the right time to review your memory settings policy, not after the global rollout.