2 min read AI-generated

ChatGPT Dreaming V3: OpenAI Rebuilds Memory From the Ground Up

Copy article as Markdown

OpenAI is rolling out the third generation of its memory system. ChatGPT now 'dreams' after every conversation, automatically updating what it knows about you. Here's what it does — and where Claude stands.

Featured image for "ChatGPT Dreaming V3: OpenAI Rebuilds Memory From the Ground Up"

ChatGPT is getting a new memory — and this time it’s more than an incremental update. OpenAI has unveiled ‘Dreaming V3,’ the third generation of its memory system. The rollout started June 4 for Plus and Pro users in the US. Free and Go users will follow over the coming weeks.

What Dreaming V3 does differently

The old system waited for you to explicitly tell ChatGPT what to remember. Dreaming V3 flips that: after every conversation, a background process automatically extracts what matters — your preferences, ongoing projects, time-sensitive context. You don’t have to actively save anything anymore.

It sounds simple, but it’s a fundamental architecture change. Instead of maintaining a list of saved facts, the system continuously synthesizes an up-to-date picture of you.

The numbers show the leap

OpenAI compared all three memory generations head-to-head:

Factual recall (remembering specific info you shared): from 41.5 percent (2024) to 67.9 percent (2025) to 82.8 percent with Dreaming V3.

Preference adherence (following dietary, behavioral, and personal constraints): from 31.4 to 55.3 to 71.3 percent.

Staying current (updating memories as circumstances change): from 9.4 to 52.2 to 75.1 percent.

The biggest jump is in staying current. Earlier versions would keep treating a past trip as ongoing long after you’d returned home. Dreaming V3 automatically refreshes such context over time.

New memory summary and controls

There’s now a memory summary page where you can see what ChatGPT has stored about you — organized by work, hobbies, travel, and preferences. You can correct details, exclude topics, or delete individual memories. If you want to switch it off entirely, you can do that in settings.

Where does Claude stand?

Claude has its own memory system, available to all users for free since March 2026. Anthropic’s approach is somewhat different — more focused on explicit memory control, with import tools for switching from other platforms. Claude’s ‘Dreaming’ feature (for Managed Agents) targets a different layer: agents that learn from their past tasks.

But the race for the best AI memory is well underway. If you use your AI daily, you benefit most from a system that thinks along and grows with you. And Dreaming V3 raises the bar noticeably.


Sources: OpenAI: ChatGPT Memory — Dreaming | gHacks: OpenAI Upgrades ChatGPT Memory With New Dreaming Architecture