ChatGPT Memory Now Updates Itself. Here Is What Changed and What to Check Today.

What You Will Learn

• What "Dreaming" does differently from the old saved-memories list
• Why time-aware memory matters for recurring work
• The one settings check to run before stale facts spread
• Who has it now and who gets it next
• How to steer what ChatGPT keeps without micromanaging it

Until last week, ChatGPT remembered things the way a sticky note does. You told it what to save, it kept a list, and that list sat there until you edited it. On June 4, OpenAI began rolling out a different approach it calls Dreaming. Now a background process watches the context that surfaces naturally in your conversations, decides what is worth keeping, and updates older memories as time passes.

The clearest example OpenAI gives is time. Tell ChatGPT you are traveling to Singapore in July, and after the trip the memory quietly revises itself to "went to Singapore in July 2026." When you are home, recommendations shift back to your home city and time zone without you saying a word. Capacity also doubled for Plus and Pro, so it forgets less of what matters.

Convenient, yes. But a memory that updates itself is also a memory you should glance at, because anything it gets wrong now propagates on its own. Here is what changed and the five-minute check worth running today.

This is a PromptHacker Premium deep dive.

The settings audit, the steer-not-micromanage approach, and the rollout timeline are below.

What Actually Changed

Three shifts matter for work. First, capture is automatic: you no longer say "remember this." Context that comes up in a normal chat is kept if it looks useful. Second, memory is time-aware: facts with an expiration, a trip, a deadline, a quarter, get revised as the calendar moves rather than going stale. Third, capacity doubled for paid users, so longer-running context survives. OpenAI reports factual recall on its own internal eval rising from 41.5 percent in 2024 to 82.8 percent in 2026.

Plain-English takeaway: you re-explain yourself less, but you should verify more, because a wrong fact now spreads without you touching it.

The Five-Minute Check to Run Today

Open ChatGPT settings, find Personalization, and open Memory. Read what it has stored about your role, your company, your projects, and your preferences. Delete anything that is wrong, outdated, or oddly specific from a one-off conversation you would not want shaping future answers. Doing this once now is far cheaper than catching a confidently wrong assumption inside a client draft next month.

Pay special attention to facts that sound right but are stale: an old job title, a former tool you have since dropped, a project that wrapped. Those are exactly the kind of thing Dreaming will keep applying until you correct it.

Steer It, Do Not Micromanage It

The point of Dreaming is to stop manually curating a list, so do not recreate that chore. Instead, correct it in plain language when it matters. If it assumes the wrong audience for your writing, say "going forward, assume my reports are for a non-technical board" and let the memory absorb it. If you are working on something you do not want remembered, a sensitive negotiation, a personal matter, use a temporary chat so nothing is stored at all.

Who Has It and Who Is Next

Dreaming began rolling out June 4 to Plus and Pro users in the United States. OpenAI says expansion to more countries and to Free and Go tiers will follow over the coming weeks, an announced direction rather than a fixed date. If you do not see it yet, you are likely in the queue.

Action Steps Summary

• Open Settings, Personalization, Memory. Read what ChatGPT currently believes about you.
• Delete anything wrong or stale. Old titles, dropped tools, finished projects.
• Correct in plain language. State your real audience and role once and let the memory absorb it.
• Use temporary chats for sensitive work. Nothing from those sessions is stored.
• Re-check after a week. Confirm the automatic capture is keeping the right things.