ChatGPT memory now updates itself between sessions

OpenAI announced an update to ChatGPT memory it calls dreaming, a background process that synthesizes context from past conversations so the next chat starts with a fresher, more relevant view of you. The first version of dreaming shipped a year earlier; this one is the same idea with more capability and a wider rollout. Concretely, OpenAI says memories now get revised as time passes rather than written once and frozen, so a note like "you are going to Singapore in July" becomes "you went to Singapore in July 2026" once the trip is over. There is also a memory summary page where users can review, add, or correct what ChatGPT thinks it knows about them, which is the first time the feature is presented with a serious surface for inspection rather than buried in settings.

The rollout pattern is the part worth tracking. The update goes to Plus and Pro users in the United States first, then more countries and Free users over the coming weeks. OpenAI attributes the move to free users to a roughly 5x reduction in the compute needed to serve dreaming, which also lets them lift memory capacity for paying users. That ratio is the substance behind the headline: a memory feature only pays for itself if the inference cost of running the background process stays below the value it adds to each next session, and OpenAI is now claiming it crossed that threshold for everyone, not just the paid tier.

Why it matters

If you use ChatGPT for ongoing work, the practical change is that memory should drift less, and you finally have a place to audit and correct it instead of guessing what got saved. If you build on the API, the more telling signal is the 5x compute reduction: it sets a floor for what an always-on personalization layer should cost, and it raises the bar for competing assistants that still treat memory as a single flat list.