What is dreaming?
Memory is what helps ChatGPT learn your preferences, projects and constraints, so future conversations start from shared context rather than from scratch. On June 4 2026 OpenAI began rolling out a more capable, scalable system for synthesizing that memory — built to tackle staleness, correctness and scale across hundreds of millions of users and multi-year time horizons.
“Dreaming” is a background process, first introduced in April 2025, that automatically curates memory by referencing your chat history — learning from many conversations and synthesizing a memory state that stays fresh and relevant, without relying on explicit requests to remember.
In practice, the old saved-memories system could feel like talking to someone who took a few notes, but still forgot everything that wasn't written down.
A background that synthesizes
Unlike saved memories — which were written only during a conversation and needed strong cues to trigger — dreaming works in the background across many conversations at once. The 2026 release is described as a significantly more capable and compute-efficient architecture built on top of dreaming.
Freshness, continuity, relevance
The stated goal is to optimize memory for three things — freshness (not stale), continuity (carries across chats) and relevance (the right context at the right time).
Reviewable in a summary
Memories synthesized by dreaming are reviewable in a memory summary page. From there you can glean the highlights of what ChatGPT knows about you, add or update information, and give instructions on what topics ChatGPT should bring up and when — and chat with the model to drill into any area.
From a notebook to a background
Memory first launched in April 2024 as “saved memories.” In April 2025, dreaming (V0) let ChatGPT reference chat context beyond the saved list — supplementing saved memories rather than replacing them. The 2026 release, Dreaming V3, is the first version capable enough to stand as the memory system on its own. See the Evolution page for the full arc.
Sources & disclaimer
This is an unofficial bilingual digest, not affiliated with OpenAI, compiled from the full text of OpenAI's June 4 2026 announcement. OpenAI presents its evaluation results as charts and publishes no citable percentage figures, so this site invents none. The Privacy & risks page is general, well-known context — not part of the announcement.
- Primary: OpenAI — “Dreaming: Better memory for a more helpful ChatGPT” (openai.com/index/chatgpt-memory-dreaming, 2026-06-04)
- For controls: OpenAI's official Memory FAQ (linked from the announcement)
- Examples quoted are OpenAI's own (underwater-camera gear, vegetarian, Singapore trip).