Think about the last thing you told ChatGPT about yourself. Not in a settings menu — just in the middle of a conversation, offhand. That you're moving to Portland in the fall. That you prefer short answers. That you're vegetarian and trying to figure out meal prep for a busy week.

ChatGPT has been storing things like that for a while. The old system mostly saved what you explicitly asked it to remember. On June 4, OpenAI shipped Dreaming V3, which changes how that memory system works at a fundamental level.

Dreaming V3 doesn't wait for you to ask. It reads across your past conversations in the background and synthesizes facts on its own. That's the change. It sounds incremental. It isn't.

How it actually works

The previous memory architecture was reactive. You told ChatGPT "remember I'm allergic to shellfish" and it wrote that down. Dreaming V3 is an asynchronous background process that runs across many conversations simultaneously — extracting facts it thinks are relevant, updating what it already knows when circumstances change, discarding what's stale.

OpenAI's own example: you mention you're going to Singapore in July. ChatGPT stores "you're going to Singapore in July." After July passes, Dreaming V3 updates that to "you went to Singapore in July 2026." Then it starts treating you as someone back home, not someone about to travel.

That's what "temporal awareness" means. Your stored memories aren't static anymore. They age. Or they're supposed to.

A roughly 5× reduction in the compute needed to serve the dreaming process is what enables the Free-tier rollout that's coming. Plus and Pro subscribers in the US got Dreaming V3 on June 4. Free and Go users, along with international rollout, follow in the coming weeks. For Plus and Pro specifically: roughly 2× more memory capacity than before, with memories that stay accurate because the system continually revises them rather than accumulating outdated context.

What it's storing — and what it isn't

Here's the part worth actually reading.

Dreaming V3 does not retain raw conversation transcripts. The system extracts structured facts — your work, preferences, travel, life circumstances — and discards the conversation text itself. Your exact words aren't retained. The conclusions it drew from your words, are.

OpenAI has also published its Memory API and invited researchers from the Berkman Klein Center and the EFF to review the system's privacy guarantees. That's more transparency than the old system ever received.

Source spread

• OpenAI — Dreaming: Better memory for a more helpful ChatGPT — [hype] — official announcement; emphasizes user control, Singapore example, temporal awareness
• TechTimes — Dreaming update limits audit trail — [skeptic] — raises concerns about reduced visibility into what the background synthesis decides to store
• Cryptonomist — How Dreaming V3 works — [builder] — technical architecture breakdown; confirms EFF and Berkman Klein audit
• Implicator — 5× compute cut enables Free-tier expansion — [builder] — compute efficiency angle and Free rollout details

Pros & cons

What's actually fine:

• Not storing raw transcripts is a meaningful choice. The old system's transcript retention behavior was ambiguous depending on session type. Dreaming V3 is cleaner: conclusions, not text.
• The transparency controls are real. There's a memory summary page showing what ChatGPT believes about you — work, preferences, travel, circumstances — and you can correct or delete any of it.
• The EFF and Berkman Klein Center review is a genuine accountability move. Not a substitute for regulation. Better than nothing.
• Temporary Chat is a clean opt-out for sensitive conversations. Nothing from a Temporary Chat session gets stored or referenced by Dreaming V3.

What deserves a side-eye:

• TechTimes flags that the background synthesis process has less auditability than the old explicit-save system. With the old system, you could see exactly what you'd asked it to remember. With Dreaming V3, the system is making inferences from your conversations — and the reasoning behind those inferences isn't surfaced in the transparency panel.
• Free users haven't gotten Dreaming V3 yet, but they will in weeks. They had less warning than paid users that a significant change to how their data is used was coming. That's a power asymmetry worth noting.
• "Temporal awareness" sounds good until ChatGPT incorrectly infers something has changed. The system is drawing conclusions from conversational context, not verified facts. It will be wrong sometimes. And you won't know until something feels off in a reply.

What to do about it

• Look at your memory settings. In ChatGPT: Settings → Personalization → Memory. Read what's listed. Correct anything wrong. This takes five minutes and is worth doing now.
• For sensitive conversations, use Temporary Chat. It's in the chat switcher at the top of the screen. Nothing from those sessions gets stored or used to update Dreaming V3.
• Watch what you share offhand. Dreaming V3 synthesizes from casual context, not just deliberate saves. If you mention something in passing — health concerns, job stress, relationship status — that may now become part of your profile.
• If you want a clean slate: Settings → Personalization → Memory → Manage → Clear all. Your memories rebuild from future conversations. A useful reset if your current profile is significantly outdated.
• For builders using the ChatGPT API: the Memory API is now published and open for external review. Worth reading if your product handles sensitive user data or has personalization features of its own.

Further reading

• OpenAI — Dreaming: Better memory for a more helpful ChatGPT — official announcement
• Cryptonomist — How Dreaming V3 works (technical breakdown)
• TechTimes — Dreaming update rewrites personalization engine, limits audit trail
• GIGAZINE — ChatGPT memory intelligently updated
• Implicator — OpenAI expands ChatGPT memory after 5× compute cut