ChatGPT's Memory System Gets a Full Overhaul: Dreaming V3 Is Live
OpenAI ships the new Dreaming V3 memory architecture — ChatGPT now "dreams" in the background to organize what it knows about you. Free access for 1 billion users for the first time, with doubled memory capacity for Plus/Pro.
ChatGPT's Memory System Gets a Full Overhaul: Dreaming V3 Is Live
OpenAI ships the new Dreaming V3 memory architecture — ChatGPT now "dreams" in the background to organize what it knows about you. Free access for 1 billion users for the first time, with doubled memory capacity for Plus/Pro.
ChatGPT's memory system just got the biggest overhaul in its history. The new Dreaming V3 architecture lets ChatGPT "dream" while you're away — automatically replaying conversation history, distilling key information, and building a continuously updated memory network. More importantly, the feature is opening up to free users for the first time.
What Is Dreaming V3
Put simply, Dreaming is a background memory-organization process. To borrow a human analogy: during the day you chat with ChatGPT, and at night it quietly "replays" your conversations, extracting, synthesizing, and updating what it knows about your preferences, habits, and needs.
This is a fundamental break from the old saved memories:
| Version | Mechanism | Result |
|---|---|---|
| saved memories (2024.4) | You had to explicitly say "remember XXX" | Like a colleague who occasionally jots down a note — anything unwritten is forgotten |
| Dreaming V1 (2025.4) | Background auto-organization of conversation history | No reminders needed, but memory quality and coverage were limited |
| Dreaming V3 (2026.6) | All-new architecture, 5x lower compute consumption | Stronger memory at lower cost, available to free users too |
Core Capabilities: Three Hard Numbers
OpenAI validated Dreaming V3 with three sets of controlled experiments.
Fact Recall: 82.8% Success Rate
On fact recall benchmarks, ChatGPT's success rate in 2026 soared to 82.8%.
A real-world example: an underwater photography enthusiast asks, "My underwater rig needs TTL strobes — what should I buy?"
- ChatGPT without memory: dumps a generic buying guide listing every TTL type, leaving the compatibility checks to you
- ChatGPT with memory: recommends specific models based on your actual kit (Sony A1 II + Nauticam NA-A1II housing + Backscatter strobe), even troubleshooting compatibility issues for you
Preference Following: 71.3% Pass Rate
The value of memory isn't just "storing information" — it's "actually using it when it matters." Dreaming V3 splits preferences into three types:
- Explicit instructions: e.g., "stop mentioning Stan"
- Personal constraints: e.g., "I'm vegetarian"
- Implicit preferences: e.g., "I live near San Francisco," so recommendations should center on the Bay Area
Say you mentioned two months ago that you love wildlife photography, demand strong hotel air conditioning, and hate noisy bars. Now you say "plan me a Singapore itinerary" — the memory-equipped version customizes directly: early-morning bird photography at the botanic gardens, a dedicated session at the bird park, quiet restaurants.
Time Awareness: Goodbye to "Stale Memories"
This may be the most practical upgrade of all. The biggest flaw of traditional memory systems: you traveled to Singapore on business in July, returned to California in September, and asked it to "find me a food delivery place that's still open" — and it recommends restaurants in Singapore.
Dreaming V3 automatically rewrites "you will go to Singapore in July" into "you were in Singapore in July 2026." It knows the trip has ended, and it knows you're back home.
What Changes for Users
Free Users
Rolling out over the coming weeks, this marks the first time the Dreaming memory feature is available to free users.
Plus / Pro Users
The new system is rolling out immediately in the US region, with memory capacity directly doubled.
Memory Management
A new "memory summary" page shows at a glance what ChatGPT remembers about you. You can add to it, edit it, or instruct it directly on which topics to bring up and when. Want to dig deeper into a particular memory? Just talk to the model. Nostalgic for old-school saved memories? One click in settings switches back.
Under the Hood: 5x Lower Compute
The most overlooked yet commercially critical number in this release: OpenAI cut the compute required to serve Dreaming by roughly 5x. It's precisely this optimization that turned "giving hundreds of millions of free users memory" from a money-loser into a viable business.
Who It's For
- Heavy ChatGPT users: with the memory upgrade, ChatGPT can finally accumulate an understanding of you over time like a true personal assistant
- Frequent business travelers: time awareness means no more correcting "where I am right now" every session
- Content creators / researchers: preference following makes recommendations increasingly precise and cuts out irrelevant noise
Toolin Editorial Team
Categories
Related articles
Hands-On with the AI Version of Alipay: Order McDonald's and Collect Energy with a Single Sentence
The AI version of Alipay has entered beta testing with an AI assistant named A Bao that operates mini programs on command; this post covers how to get an invitation code plus the hands-on experience.
Major Claude Design Update: One-Click Design System Import and Two-Way Code Sync
Anthropic has shipped a major Claude Design update with design system import, two-way /design-sync and /design code sync, and one-click export to 9 platforms.
GLM-5.2 Open Source, Put to the Test: On the Same Level as Opus 4.8 and GPT 5.5
GLM-5.2 is the first Chinese open-source model to enter the new "Big Three", with 1M long context and coding ability close to closed-source flagship level, MIT-licensed and ready to use.
Kimi Work Adds Goal Mode and a Plugin Center, with 50% Off Quota in June
Kimi Work launches Goal Mode with 24-hour continuous autonomous operation and a new Plugin Center supporting Baidu Netdisk, DingTalk, Feishu, and other apps, with all task quota consumption halved in June.
Alibaba's HappyOyster 1.0: Create a World You Can Walk Into with a Single Sentence
Alibaba has released the world model product HappyOyster 1.0, which generates open worlds with real-time exploration and physics interaction from a single sentence, offering World Exploration and Real-Time Directing modes; the API is expected to open in early July.
GLM-5.2 Released as Open Source: 1M Context, No.1 Worldwide for AI Coding
Zhipu has released GLM-5.2 with a million-token context window, ranked first among globally available models on Code Arena, open-sourced under the MIT license so developers can deploy and use it commercially at will.