OpenAI Gives ChatGPT ‘Dreaming’: Memory That Updates Itself in the Background

OpenAI on June 4 announced a major overhaul of ChatGPT's memory it calls "Dreaming" — a background process that continuously learns from a user's past conversations and synthesizes a single, up-to-date picture of who they are and what they care about. Instead of asking people to explicitly save facts, Dreaming captures context that surfaces naturally in conversation and becomes the standalone foundation for ChatGPT's recall.

The biggest shift is that memory now keeps itself current. OpenAI's own example: a note that "the user is going to Singapore in July" automatically rewrites itself to "the user went to Singapore in July 2026" once the trip has passed — with no action from the user. Time-bound references like "next Saturday" age correctly rather than lingering as stale facts, and preferences around things like diet, location, cameras or ongoing projects are inferred from context instead of having to be restated.

That is a clean break from how ChatGPT memory worked before. The original 2024 feature behaved "just like a notebook," requiring explicit commands such as "remember that I am a vegetarian," and a 2025 iteration added broader context awareness but stayed supplementary. Dreaming instead runs transparently in the background and continuously updates the synthesized memory state, aiming to surface the freshest and most relevant context for each new chat.

OpenAI says the feature remains optional and controllable: users get a memory summary page that shows the synthesized information, and they can edit or remove specific items. On availability, Dreaming began rolling out to ChatGPT Plus and Pro users in the US, with expansion to additional countries and to Free and Go tiers over the coming weeks — a wider rollout the company attributes to cutting the compute required to serve Dreaming by roughly 5x. Details were published in OpenAI's announcement, "Dreaming: Better memory for a more helpful ChatGPT."

Dreaming signals where consumer assistants are heading: away from manual, notebook-style memory and toward systems that quietly build and maintain a long-term model of the user. The upside is a more genuinely personalized assistant; the cost is a sharper version of the perennial questions about transparency and control over what these systems infer and retain. That tension is why the editable memory summary may end up mattering as much as the synthesis running behind it.