OpenAI's new memory system is interesting because it moves the awkward part of personalisation into the background. The old saved-memory model made ChatGPT feel like a notebook with a chat box attached. Dreaming V3 points at something more unsettled: a system that keeps revising its idea of the user while the user is still changing.

The company [announced Dreaming V3][dreaming] on 4 June, describing it as a more capable and scalable architecture for synthesising memory in ChatGPT. The rollout starts with Plus and Pro users in the United States, with additional countries and Free and Go users due over the following weeks. In the release notes, OpenAI gives the plain-language version: memory should stay more up to date, reduce stale or contradictory saved memories, and let people review what the system thinks it knows through memory sources or a memory summary.

This is not just a quality-of-life change. It changes the implied contract of the interface. Saved memories, launched in 2024, were legible because they looked like entries: facts the user could ask the model to keep. In 2025, OpenAI added the first version of dreaming, which let ChatGPT draw on chat history in the background. Dreaming V3 makes that background process more central, more compute-efficient, and more visibly managerial. OpenAI says the compute needed to serve dreaming has fallen by about five times.

There is a clean product reason for all of this. A model with better memory can stop wasting the first third of every useful conversation on reintroductions. It can know the camera gear, the food preferences, the long-running project, the thing you corrected last month and don't want to correct again. In that sense, memory is less like storage than grip. The system holds the shape of the work for longer.

However, the word "memory" does a lot of softening here. We use it because it sounds intimate and human, but the machinery is closer to ongoing profile maintenance. The official FAQ says the memory summary updates automatically as you chat, and that memory sources can include past chats, saved memories, custom instructions, files, and Gmail depending on the plan. That may be useful. It is also a reminder that personalisation is made from access.

I don't mean this as a simple objection. The dumb version of privacy talk says all memory is creepy and all forgetting is freedom. That is too neat. Most people who use these systems for real work want continuity, and continuity requires some kind of retained context. I wrote recently about MiniMax making long context feel cheaper, but memory is a different pressure. Context is what the model can hold in the room. Memory is what the product decides to bring back into the room next time.

That distinction matters because it shifts agency around. With a long context window, the user often chooses what to paste. With memory, the system helps choose what follows them. OpenAI is trying to make that visible through summaries and sources, which is the right direction, but visibility is not the same thing as authorship. A summary can be corrected and still leave you with the strange feeling that another system has been drafting a version of you in parallel.

There is an old AI lesson hiding inside the new surface. ELIZA worked because people supplied more continuity than the program deserved. ChatGPT has the opposite problem now: the continuity is increasingly real, operational, and hard to dismiss as projection. The question is not whether the model remembers. The question is who gets to edit the memory before it starts editing the conversation.

Sources: