OpenAI overhauls ChatGPT memory with ‘dreaming’

TL;DR:

  • OpenAI is rolling out a more capable ChatGPT memory architecture built on a background process it calls “dreaming”.
  • The system synthesises memories across many conversations to keep context fresh, correct and relevant over years.
  • It launches first for US Plus and Pro users, expanding to more countries and free tiers over the coming weeks.

OpenAI is reworking how ChatGPT remembers, aiming to fix the staleness and scaling problems that surface when memory must serve hundreds of millions of users over multi-year horizons. The new architecture, dubbed Dreaming V3, builds on a background curation process the company first introduced in 2025. Rather than relying on explicit “remember this” instructions, it synthesises context from many past conversations so future chats start from shared knowledge of a user’s preferences, projects and constraints.

From note-taking to synthesis

OpenAI frames the evolution in three stages: 2024’s “saved memories”, which it likens to someone who jotted a few notes but forgot everything unwritten; 2025’s first “dreaming” layer, which began curating memory automatically in the background; and now a more compute-efficient system designed to carry context forward, follow stated preferences and stay current as time passes. A memory summary page lets users see what the model has inferred, edit it, and steer which topics it raises. The capability is reviewable, the company stresses — an attempt to pair deeper personalisation with user control.

For UK readers, the update is a capability note worth filing rather than a UK-specific story, but the privacy implications travel. A system that quietly synthesises a long-running profile from conversations raises familiar questions about consent, data retention and transparency under UK GDPR — concerns regulators have pressed on AI personalisation before. It also lands as ChatGPT’s reach keeps climbing, having recently passed one billion monthly users.

Looking forward

The rollout starts narrow — US paid tiers first — before reaching free and international users in the weeks ahead. As memory becomes central to how assistants “know” their users, the balance between helpful continuity and surveillance-by-default will be the line OpenAI, and its regulators, have to walk.

Share this article