OpenAI starts rolling out Dreaming V3, a new memory synthesis system for ChatGPT

OpenAI began rolling out a new long-term memory architecture for ChatGPT on June 4, 2026. The system, which the company calls Dreaming V3, replaces the saved-memories plus Dreaming V0 combination that has powered ChatGPT's cross-conversation context since April 2025. Plus and Pro users in the United States get the update first, with Free and Go users and additional countries to follow over the coming weeks.

What's new

• A new background process that synthesizes memory from a user's chat history, designed to address what OpenAI describes as the "staleness, correctness, and scalability challenges" that surface when memory runs across hundreds of millions of users and multi-year time horizons.
• A visible memory summary page where users can review what ChatGPT has learned about them, add or update facts, and write instructions about which topics ChatGPT should bring up and when.
• An internal version generation OpenAI labels Dreaming V3 (2026), succeeding the Dreaming V0 background process introduced in April 2025 and the original saved memories feature launched in April 2024.
• Initial availability is limited to ChatGPT Plus and Pro in the U.S., with Free and Go tiers and other regions rolling in "over the coming weeks."

Context

Memory in ChatGPT started narrow. The April 2024 saved-memories feature only wrote new entries when a user gave an explicit cue such as "remember I'm traveling to Singapore in July," and OpenAI itself now describes that experience as "talking to someone who took a few notes, but still forgot everything that wasn't written down."

Dreaming V0 in April 2025 added a background process that curated memories automatically by referencing chat history, which OpenAI says produced a "step-function improvement in ChatGPT's ability to personalize responses and offset the staleness of saved memories." The company is candid that V0 "never was sufficient as a standalone memory system" — the V3 launch promotes dreaming from a supplement to the primary memory layer, on top of "a significantly more capable and compute-efficient memory architecture."

Three evaluation goals appear repeatedly in the announcement: carrying forward useful context, following preferences and constraints (the example given is a vegetarian user), and staying current over time (a planned birthday party becoming a past event once Saturday turns into Sunday). The post benchmarks the three generations — 2024 saved memories, 2025 saved memories plus Dreaming V0, and 2026 Dreaming V3 — against those three objectives.

Why it matters

Memory is the part of ChatGPT a user cannot replace by switching providers. A new model from a competitor can match capability on day one; multi-year accumulated context cannot be moved. That makes the memory layer the most durable retention surface OpenAI owns, and changing the architecture is therefore high-stakes.

Two things stand out about how OpenAI framed this release. First, it is explicitly a scalability rebuild — the language about hundreds of millions of users and multi-year horizons signals that the prior architecture had compute and freshness limits the company was unwilling to keep paying. Second, the memory summary page is the user-visible counterweight: opaque background curation paired with a single page where users can edit, delete, or instruct.

The staged rollout is also worth flagging. By starting with Plus and Pro users in the U.S. only, OpenAI keeps the blast radius small if the new synthesis layer has regressions on long, complex memory states. Free and Go users and non-U.S. regions inherit a more battle-tested version. Anyone running ChatGPT memory in a production-adjacent workflow should expect noticeable behavior changes in what is remembered and how it is surfaced over the next several weeks.