OpenAI started rolling out a significant change to how ChatGPT remembers things on June 4, and most users likely won’t notice it happened. The update, called Dreaming V3, quietly replaces the old saved-memories list with a background synthesis process that builds a model of who you are from your entire conversation history.

That sounds like a minor product update. For businesses integrating ChatGPT into how their teams work, it’s worth paying attention to.

What Actually Changed

The previous memory system in ChatGPT worked the way most people expected: ChatGPT would occasionally suggest “should I remember this?” or you could explicitly instruct it to save something. The result was a human-curated list you could review, edit, and delete.

Dreaming V3 works differently. After conversations end, a background process runs that reads across your full conversation history and updates what ChatGPT knows about you automatically. You didn’t ask it to remember anything. It just synthesizes patterns from what you’ve said.

OpenAI describes three user-facing controls added alongside this: a memory summary page showing what ChatGPT has synthesized, controls to manually add or update remembered details, and topic preferences for what ChatGPT should and shouldn’t bring into conversations. Plus and Pro subscribers get twice the memory capacity.

The rollout is currently live for Plus and Pro users in the United States, with Free users and enterprise customers coming in the weeks ahead.

Where the Business Concern Comes In

The explicit memory list had a useful property for enterprise IT teams: it was auditable. You could see exactly what information was being used to personalize responses. That visibility made it easier to reason about what ChatGPT knew, and to clear it before discussing something sensitive.

The background synthesis approach is more opaque. ChatGPT is drawing inferences from years of conversation history, not a list that was deliberately created. What it synthesizes about a user’s role, priorities, clients, or habits isn’t directly visible in the way a saved-memories list was.

OpenAI has indicated that enterprise deployment of Dreaming V3 will come with specific privacy controls, but the details of those controls aren’t fully public yet. For companies that have already deployed ChatGPT Enterprise, it’s worth getting ahead of this before it rolls out to your user base.

What This Means for Business

Context retention is genuinely better. One consistent frustration with AI tools is that they forget everything between sessions. Dreaming V3 addresses that in a meaningful way. A data analyst who regularly runs the same types of reports, or a consultant who works within a specific industry, will find ChatGPT more useful because it learns their context over time rather than starting fresh every conversation.

The governance question has shifted. When memory was explicit and user-controlled, the governance question was simple: who can see the saved list? Now the question is more nuanced: what has the model synthesized, what triggered that synthesis, and how does it affect responses? Enterprise IT teams should start building answers to those questions now, not after rollout.

Data classification still applies. Regardless of how memory works, the guidance for your teams should be consistent: don’t share client-identifiable information, proprietary data, or anything you wouldn’t want stored in a vendor’s cloud with ChatGPT. Dreaming V3 makes that guidance more important, not less, because the model is now synthesizing patterns across conversations rather than storing discrete facts.

Free users will have this soon. If your teams include people using the free tier of ChatGPT for work tasks, the same dynamics apply. OpenAI has significantly reduced the compute cost of running Dreaming (reportedly 5x more efficient than previous versions), which is why they can afford to extend it to free users. That means the memory question isn’t just a premium product concern.

The Bigger Pattern

This update reflects something broader happening in AI tools right now: personalization is moving from explicit configuration to autonomous inference. That’s better for casual use. For enterprise use, it creates a category of AI governance question that most companies haven’t fully thought through yet.

The companies that will handle this well aren’t necessarily the ones who restrict AI use. They’re the ones who invest in understanding how these tools actually work, train their teams accordingly, and build policies that are specific enough to be useful rather than generic enough to be ignored.

Enterprise DNA works with businesses at exactly this stage of AI adoption. If your team is using ChatGPT and you haven’t reviewed your AI data policies recently, this update is a good prompt to do that.