OpenAI's 'Dreaming' memory system goes mainstream in ChatGPT

OpenAI began rolling out a rebuilt ChatGPT memory architecture on June 4, 2026, built around a background process it calls "dreaming" — a system that automatically curates and synthesizes memories from chat history rather than relying on user-issued commands like "remember I'm traveling to Singapore in July."

The update is live today for Plus and Pro users in the US, with a rollout to additional countries and to Free and Go users planned over the coming weeks. The enabler, according to OpenAI, is engineering: recent improvements cut the compute required to serve dreaming to Free users by approximately 5x, finally making the system "practical to serve at scale."

The stakes are straightforward. ChatGPT memory now spans hundreds of millions of users and multi-year time horizons, and OpenAI says the previous architecture suffered from staleness, correctness, and scalability problems at that scale. Persistent memory is also the mechanism that turns ChatGPT from a stateless chatbot into an assistant that, in OpenAI's framing, can "know" you and "do more for you" — a competitive frontier where Google's Gemini and other assistants are pushing equally hard.

From saved memories to dreaming

Memory first shipped in April 2024 as "saved memories." The feature wrote memories only during conversations and triggered on strong, explicit cues. OpenAI's own assessment of that first generation is blunt: interacting with it "could feel like talking to someone who took a few notes, but still forgot everything that wasn't written down." Saved memories also "tend to go stale over time and eventually become incorrect or irrelevant."

In April 2025, OpenAI introduced the first version of dreaming, giving ChatGPT the ability to reference chat context beyond the saved memories list and to curate memory automatically in the background. Over the past year, dreaming supplemented saved memories and produced what OpenAI calls "a step-function improvement" in personalization — but the company says it "historically was never sufficient as a standalone memory system."

Today's launch changes that. OpenAI describes the new architecture as "significantly more capable and compute-efficient" and built on top of dreaming. In its internal taxonomy, the progression runs from saved memories (2024), to saved memories plus Dreaming V0 (2025), to Dreaming V3 (2026).

A visible memory summary

The synthesized memories are exposed through a new memory summary page, where users can review what ChatGPT knows about them, add or update information, and instruct ChatGPT on which topics to raise and when. Users who want more detail on a specific area can simply chat with the model. OpenAI also points users to its Memory FAQ for release details and user controls.

How OpenAI measures memory

OpenAI evaluates memory against three objectives: carrying forward useful context (you tell ChatGPT something once and it persists), following preferences and constraints (say you're vegetarian, and future dining suggestions respect it), and staying current over time — a criterion OpenAI illustrates with the line: "Memory should account for the passage of time. Imagine 'The user is planning their birthday party for next Saturday'; eventually, Sunday arrives."

The company published side-by-side examples for each criterion, contrasting generic responses with memory-driven ones.

Context carry-forward. Asked what to buy for TTL flash in an underwater photography setup, the memory-less model produces a generic compatibility checklist covering strobes, triggers, and cables, leaving the user to work out the parts. With memory, the model recalls the user's actual gear — a Sony A1 II in a Nauticam NA-A1II housing with Backscatter Mini Flash 3 and Inon Z-330 strobes — and recommends a specific trigger (Backscatter Smart Control TTL LED Nauticam Flash Trigger for Sony, SKU BS-TR-SN2), while warning that no single trigger delivers true TTL for both strobe systems.

Preference following. Asked to plan a few free days in Singapore after a July work trip, the memory-less model returns a standard tourist itinerary. The memory-equipped model builds the plan around the user's known constraints — a wildlife photography bias, an unusually strict air-conditioning requirement for sleep, and a preference for quiet dinners over crowded bars — ranking Bird Paradise and the Night Safari high and flagging that "the hard gate is still AC" when it comes to hotels. OpenAI notes that preferences can be explicit instructions ("don't bring up Stan again"), personal constraints ("I'm vegetarian"), or implicit signals ("I live near San Francisco" should shape local recommendations).

Freshness over time. This is where dreaming's automatic updates matter most. OpenAI frames the problem plainly: "Time doesn't stop when your chat ends." A user who once asked for dinner recommendations in Singapore would, under a stale-memory system, keep getting Singapore suggestions after the trip ended. In OpenAI's example, the stale-memory model assumes the user is still in Singapore at 5:19 AM Sunday and lists 24-hour prata and McDelivery options. The updated dreaming system revises "You're going to Singapore in July" to "You went to Singapore in July 2026" once the trip ends, and correctly grounds the same take-out request in the user's home area near Portola Valley and Ladera, California, suggesting the Alpine Inn and Taverna Portola Valley. OpenAI says dreaming delivers "a substantial lift" on its time-sensitivity evaluations, and that the new system improves fact recall in its context evals as well.

Why the 5x compute cut matters

For over a year, dreaming-based memory existed only for paying users because serving it to free users was too expensive. The roughly 5x reduction in serving compute changes the economics, and OpenAI says it will both extend dreaming to Free users and increase memory capacity for Plus and Pro subscribers.

That consolidation gives OpenAI what it calls "a shared memory foundation for all users" — a single architecture rather than a tiered split between a premium memory system and a basic one. OpenAI calls this update "our most capable memory system yet" and says it will keep improving it.

The move also signals where the memory race is heading: the differentiator is no longer whether an assistant remembers, but whether it forgets correctly — updating, expiring, and re-ranking what it knows as users' lives change. If Dreaming V3 delivers on the freshness evals OpenAI describes, expect the company's rivals to answer with their own background memory synthesis rather than more saved-notes lists.