Dreaming: Better memory for a more helpful ChatGPT

ChatGPT's memory grows up

OpenAI is rolling out what it calls its most capable memory architecture yet for ChatGPT - an upgrade aimed squarely at the three problems that plague memory at scale: it goes stale, it gets things wrong, and it's expensive to run across hundreds of millions of users and multi-year time horizons.

From sticky notes to "dreaming"

The feature has evolved in stages, and the framing is useful:

• Saved memories (April 2024) only recorded what you explicitly asked it to remember, which in practice felt like talking to someone who jotted a few notes and forgot everything else - and those notes went stale over time.


• Dreaming (first introduced April 2025) added a background process that automatically curates memory by drawing on your chat history, so context that comes up naturally gets retained without an explicit "remember this."


• The new release is a more capable, compute-efficient architecture built on top of dreaming - referred to as Dreaming V3 - that finally stands on its own rather than just supplementing saved memories.

Crucially, what the system synthesizes is reviewable: a memory summary page lets you see the highlights of what ChatGPT knows, correct or add details, and tell it which topics to raise and when.

What "good memory" is supposed to do

OpenAI frames quality around three objectives, each of which it says the new system improves on internal evals:

• Carry context forward so you don't reintroduce yourself every chat - for instance, asking for camera gear compatible with "my setup" and getting answers tuned to gear you discussed months ago.


• Follow preferences and constraints, whether explicit ("I'm vegetarian"), instructional ("don't bring that up again"), or implicit ("I live near San Francisco").


• Stay current over time, so "I'm going to Singapore in July" becomes "went to Singapore in July 2026" once the trip passes, and recommendations snap back to your home location.

Why the scale story matters

The headline operational win is cost: recent improvements cut the compute needed to serve dreaming to Free users by roughly 5x, which is what makes it practical to extend to everyone and to raise memory capacity for paying tiers. The bigger point is strategic - dreaming now gives OpenAI a shared memory foundation across all users, and personalization that deepens over time is central to making ChatGPT stickier and more useful. Memory controls and an FAQ accompany the release.