ChatGPT's New Memory Now Updates Itself

OpenAI has given ChatGPT the closest thing yet to a long-term memory. On 4 June the company rolled out a system it calls Dreaming V3, which reads across years of a user's past conversations and builds a running picture of who they are, rather than storing a flat list of facts the way the old memory feature did.

The headline trick is that memories now maintain themselves. A note reading 'you're going to Singapore in July' rewrites itself to 'you went to Singapore in July 2026' once the trip has passed, with no prompting from the user. OpenAI says context that comes up naturally in conversation is captured on its own, so people no longer have to tell the assistant what is worth keeping.

Better recall, cheaper to run

OpenAI reports that factual recall on its own internal evaluation climbed from 41.5 percent in 2024 to 82.8 percent in 2026. Just as important for reach, the company says it cut the compute needed to serve the feature by roughly five times, which is what makes a free-tier rollout practical. For now Dreaming V3 is reaching Plus and Pro subscribers in the United States first, with wider tiers and international markets to follow.

The trade-off

An assistant that remembers more is also one that watches more closely. OpenAI is pairing the launch with controls that let people decide how much the system retains, though Tech Times notes the new design also limits the audit trail, making it harder to see exactly why the model surfaced a particular memory. That is the tension worth watching. Persistent memory is what turns a chatbot from a clever stranger into something that feels like it knows you, and the same feature that makes it useful is the one that asks you to trust it with more.

It also raises the stakes on getting memory right. A model that quietly carries forward a wrong assumption about you, and acts on it across months of conversations, is a different kind of problem from one that simply forgets.