OpenAI's dreaming memory system moves from a notepad to background synthesis

From saved notes to a background curator

OpenAI is shipping a memory architecture it calls Dreaming V3, available to Plus and Pro users in the US first, with Free and Go users and additional countries following over the coming weeks. The company frames it as a fix for three specific failure modes it observed at scale: staleness, correctness, and the cost of serving memory across hundreds of millions of users over multi-year horizons.

The lineage matters here. Saved memories launched in April 2024 and only wrote information when a strong cue triggered it — an explicit "remember I'm traveling to Singapore in July." OpenAI's own description 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."

Dreaming, introduced in April 2025, changed the mechanism. Instead of writing during a conversation, it runs a background process that references chat history and synthesizes a memory state, picking up context that surfaces naturally without an explicit request to remember it. The 2026 release makes that synthesis the foundation rather than a supplement to saved memories.

The summary page is the actual control surface

For a system that learns from many conversations without being asked, the frontend problem is legibility: if the model is quietly building a model of you, you need to see it. OpenAI's answer is the memory summary page, where dreaming-synthesized memories are made visible.

The page is described as more than a viewer. You can glean the highlights of what ChatGPT knows, add or update information about yourself, and give instructions on which topics ChatGPT should raise and when. For anything you want to dig into, the fallback is to just chat with the model. That combination — a scannable summary plus conversational drill-down — is the interaction pattern OpenAI is betting on for making an opaque background process feel accountable.

It also reframes preferences as first-class, editable objects. OpenAI splits them into response instructions ("don't bring up Stan again"), stated constraints ("I'm vegetarian"), and implicit signals ("I live near San Francisco" shaping what counts as local). The summary page is where a user can correct any of these before they compound across future chats.

Three objectives, shown through worked examples

OpenAI evaluates memory against three goals and illustrates each with a side-by-side. Carrying context: a photography question where the model without memory returns a generic TTL compatibility checklist, while the memory version recalls a specific rig — Sony A1 II in a Nauticam housing with a Backscatter Mini Flash 3 and Inon Z-330 strobes — and narrows to concrete SKUs.

Following preferences: a Singapore trip where the memory-enabled response is built around known constraints like wildlife photography, strong AC for sleep, and quiet dinners over crowded bars, rather than a touristy highlights list.

Staying current is the most technically pointed of the three. OpenAI's example shows a stale system still placing the user in Singapore at 5:19 AM local time, while the updated system recognizes the trip has ended and answers from the user's home area near Portola Valley. As the company puts it, dreaming revises "You're going to Singapore in July" to "You went to Singapore in July 2026" once the trip passes — memory that accounts for the passage of time rather than freezing at the moment it was written.

A 5x compute cut is what unlocks Free users

The scalability claim is the least flashy but most consequential line in the release. OpenAI says recent work reduced the compute required to serve dreaming to Free users by roughly 5x, and that this — not just quality — is what finally makes a broad rollout practical.

While dreaming-based memory has been available to Plus and Pro users for some time, we are only now able to offer Free users a version that meets our quality bar and is practical to serve at scale.Montana Labs

That admission is worth noting: a background synthesis process is expensive precisely because it runs continuously across a user's history, not only during live turns. The same efficiency gain also funds higher memory capacity for paying tiers, so the cost work benefits both ends.

What a shared memory foundation commits OpenAI to

OpenAI describes the result as "a shared memory foundation for all users" — one architecture underneath Free, Plus, and Pro rather than a premium feature bolted onto a stripped-down base. That is a real product commitment. It means the summary page, the preference categories, and the time-aware revision behavior all have to hold up as the standard experience, not an enthusiast setting.

The implication for anyone building on ChatGPT is that persistent, silently-updated user context is becoming the default assumption behind every conversation. The interesting design question is no longer whether the model remembers, but whether the editable summary keeps that memory correctable at the pace it accumulates — because a system that revises itself in the background is only trustworthy if the front door to inspect and fix it stays open.