OpenAI rolls out 'Dreaming V3' memory and extends smarter context to free users

OpenAI started shipping Dreaming V3, a ground-up rebuild of ChatGPT's memory system that automatically synthesizes relevant context from a user's prior conversations rather than relying on manual 'remember this' instructions. The feature went live for Plus and Pro users in the US on June 4, with free and Go tiers slated to follow within weeks, per OpenAI's own messaging and 9to5Mac.

The mechanism shifts memory from an opt-in scratchpad to an always-on synthesis layer: the model builds a working model of the user's preferences, projects, and history and injects it as needed. Developers on dev.to and TechTimes praised the automatic context synthesis but raised pointed concerns about audit-trail transparency and whether users get clean opt-out controls over memory training.

Separately on June 4, Sam Altman met with lawmakers to argue against requiring government approval before releasing new models — a notable contrast with Anthropic's call the same week for a coordinated development pause. The split captures the week's policy fault line: Anthropic leaning toward coordination/caution, OpenAI toward shipping speed and regulatory restraint.

What to watch: privacy regulators' response to always-on memory, enterprise controls for Business accounts, and whether 'Dreaming V3' meaningfully improves task continuity versus prior memory. The audit-trail question is the sleeper risk — synthesized memory that users can't inspect is a support and compliance liability.