ChatGPT now 'thinks' after the conversation: Dreaming V3 and the memory debate

On June 4, 2026, OpenAI announced Dreaming V3, an update that overhauls ChatGPT's memory system. The name is a touch dramatic, but the idea is concrete: ChatGPT now runs in the background after a conversation ends and digests "what mattered" on its own. It's rolling out first to Plus and Pro users in the US, then to Free and Go tiers within a few weeks.

What was announced

The core of Dreaming V3 is a background process that runs after a conversation ends: the model distills what was meaningful in that session and writes it into long-term memory. OpenAI says this cuts the compute needed for memory synthesis by roughly 5x — cheaper, and able to run more often. Memory storage also doubles for Plus and Pro users.

The same update brings a Lockdown Mode: a security layer that lets enterprise admins restrict capabilities — web browsing, research mode, agent mode, file downloads — on a per-role basis. So while OpenAI strengthens memory, it also ships tools to constrain it.

The contentious part: whose decision is memory?

The privacy angle matters more than the headline. An arXiv study published in February 2026 found that 96% of sampled memory entries (2,050 entries from 80 users) were created unilaterally by the system, without asking the user. In other words, the model largely decides for itself what to remember.

Since Dreaming V3 makes this memory stronger and more automatic, the question gets bigger: what is the AI accumulating about you, and do you choose it, or does it? I wrote about AI baked into the browser widening the data-collection surface in the piece on browser-native AI; persistent memory is the assistant-side version of the same worry.

For GDPR (and Turkey's KVKK) this is not trivial: the principles of "data minimization" and "explicit consent" sit in tension with a system that accumulates data on your behalf. If you build a product and pass user data to assistants like this, knowing what that automatic memory stores is your responsibility.

What it means for a solo builder

For personal use, strong memory is a comfort: you don't have to re-supply context over and over, the assistant "remembers" your project. When I separated the agent from the assistant, I said the assistant's value lies in continuity; Dreaming V3 deepens exactly that continuity.

But a caveat: think twice before writing sensitive information (customer data, an API key, an unannounced product plan) into an assistant with memory on. Once something is "remembered," it leaves your control and enters the model's long-term context. Learning your memory settings and constraint tools like Lockdown is no longer optional.

Conclusion

Dreaming V3 is one step in the trend moving assistants from tools that "start from zero every time" to tools that "know you." The comfort is real, but so is the cost: the more it remembers, the more it accumulates. The good news is that OpenAI ships brakes like Lockdown in the same package. The real work is knowing how to use them.

As a solo builder, I both benefit from memory and feel uneasy about it. It's great that the assistant remembers my project; but "I didn't choose what it remembers" doesn't sit easy with me. So I never write anything sensitive into a window with memory on — an old-fashioned habit, but it protects me.