Memory used to be something you asked ChatGPT to do. Dreaming V3 removes that step.
ChatGPT Dreaming V3 launched on June 4, 2026, initially for Plus and Pro subscribers in the United States, according to OpenAI’s announcement. Free and Go tier users, along with international markets, will follow in subsequent rollout waves per OpenAI, timing unspecified. The launch date is corroborated by the prior coverage in this hub from June 4.
The part nobody mentions in the feature announcements: this isn’t memory in the sense of saved notes. OpenAI describes Dreaming V3 as a continuous background synthesis process that captures and updates user context without requiring explicit commands. The previous model required a user to say “remember this.” Dreaming V3 doesn’t. According to OpenAI, the system also automatically adjusts how it represents past events as time passes, a future travel date becomes a past event in the model’s representation once the date has occurred. That’s a temporal self-updating mechanism, not just persistent storage.
The catch is that autonomous synthesis changes what enterprise acceptable-use policies actually cover. Most policies were written assuming memory was user-initiated. A system that continuously synthesizes context from conversations, without an explicit save action, fits differently into data minimization frameworks, retention policies, and employee monitoring regulations. Legal teams haven’t had time to catch up to the launch timeline.
OpenAI says users can view, manually edit, and delete synthesized memories via a centralized memory summary page. According to OpenAI, opting out of model training on chats and memories is available under Settings → Data Controls. These are vendor-described controls, the UI exists, but complexity of the opt-out path matters as much as availability of the opt-out.
For context: this follows a visible pattern. Google’s Gemini Spark Beta introduced always-on cloud-VM agent behavior in late May. Dreaming V3 brings a parallel capability to the dominant consumer AI platform. Memory is transitioning from a discrete user-controlled feature to a platform infrastructure primitive, a change happening across multiple vendors simultaneously, not just at OpenAI.
Enterprise teams deploying ChatGPT under a business agreement need to assess two things. First, whether Dreaming V3 is active under their current contract tier and what controls the enterprise agreement provides beyond the consumer opt-out. Second, whether existing acceptable-use policies explicitly address AI systems that synthesize context autonomously rather than on user direction. Most don’t.
The API isn’t affected, this is a consumer ChatGPT interface change only, per OpenAI. Teams using the API directly won’t see this behavior in their applications. That boundary is important for developers who might otherwise assume the memory architecture changes apply to their API-integrated products.
Wait for OpenAI’s enterprise documentation update before treating the consumer opt-out as sufficient for business deployments. Enterprise controls and consumer settings aren’t always the same product.