How is Dreaming V3 different from the old ChatGPT memory?

Old memory was a static list of facts ChatGPT chose to remember — they stayed until you manually deleted them. Dreaming is dynamic: memories are automatically revised as time passes. For example, 'You're going to Singapore in July' becomes 'You went to Singapore in July 2026' after the trip date passes. Dreaming also unifies four memory sources (saved memories, past chats, files, app connectors) into one retrieval system.

Does Dreaming work in the background?

Yes. Dreaming processes memories in the background, even when you're not actively chatting with ChatGPT. It groups related facts, resolves contradictions, and compresses redundant information continuously. This means your context is always fresh when you start a new conversation.

Can I control what ChatGPT dreams about me?

Yes. Open Settings → Personalization → Memory. You can view a readable summary of everything ChatGPT remembers, edit or delete individual facts, turn off memory globally, or use Temporary Chat mode. Starting June 2026, you can also set per-domain reference rules (e.g., 'don't use health memories during work chats').

When will Dreaming V3 be available to all users?

Dreaming V3 rollout began June 4, 2026 for Plus and Pro users in the United States. OpenAI's June 2026 update expanded to Free and Go users globally. The full rollout is expected to complete by mid-July 2026.

Does Dreaming use more compute resources?

OpenAI stated at rollout that recent improvements significantly reduced the computational resources required for Dreaming, making widespread availability feasible. The background processing runs on OpenAI's inference infrastructure and doesn't affect ChatGPT's response speed during active conversations.

Quick Answer

What Is ChatGPT Dreaming V3? OpenAI's Memory Architecture Explained

Q: What is ChatGPT Dreaming V3?

ChatGPT Dreaming V3 is OpenAI's new memory architecture rolled out starting June 4, 2026. Unlike the previous static memory system, Dreaming automatically curates and synthesizes memories from past conversations, updates them as circumstances change, and pulls context from saved memories, past chats, uploaded files, and connected apps. It runs as a background process, even when users are offline.

Published: June 28, 2026

What Is ChatGPT Dreaming V3? OpenAI’s Memory Architecture Explained

OpenAI’s “Dreaming” memory upgrade, rolling out from June 4, 2026, is the most significant change to how ChatGPT personalizes responses since memory was first introduced. It moves from a static, manually-managed memory system to a dynamic, auto-synthesizing architecture that continuously updates and refines what ChatGPT knows about each user.

Last verified: June 28, 2026

The short definition

ChatGPT Dreaming V3 is OpenAI’s new memory architecture that continuously and automatically curates, updates, and synthesizes personalized context from all user interactions — without requiring manual memory management.

Think of it as the difference between a contacts list you have to update manually (old memory) and an address book that automatically updates contacts from your emails, messages, and calendar events (Dreaming).

The architecture

Four memory sources, one retrieval system

Previous versions of ChatGPT memory treated each source separately. Dreaming unifies them:

Saved memories: Facts ChatGPT explicitly remembers (existing feature)
Past conversations: Context from all chat history
Uploaded files: Information from documents, images, and data you’ve uploaded
Connected apps: Data from Gmail, Google Drive, and other connectors

All four sources are synthesized into a single memory context that ChatGPT draws from during conversations.

Dynamic memory updates

The key innovation: memories are time-aware and self-updating.

Before Dreaming (static):

User: “I’m planning a trip to Singapore in July” ChatGPT saves: “User is planning a trip to Singapore in July” August arrives ChatGPT still thinks: “User is planning a trip to Singapore in July”

After Dreaming (dynamic):

User: “I’m planning a trip to Singapore in July” ChatGPT saves: “User is planning a trip to Singapore in July” August arrives Dreaming updates: “User went to Singapore in July 2026”

Background processing

Dreaming runs as a continuous background process, not triggered by conversations:

Groups related facts into coherent summaries
Resolves contradictions between old and new information
Compresses redundant memories
Operates even when the user is offline
Scales to hundreds of millions of users across multi-year timeframes

Why the name “Dreaming”

OpenAI chose “Dreaming” as a metaphor for how the system works — like how human brains consolidate memories and learn from past experiences during sleep, ChatGPT’s Dreaming system continuously processes and integrates past interactions to build a richer understanding of each user.

The “V3” designation indicates this is the third major iteration of the memory system:

V1 (2024): Manual memory — users explicitly told ChatGPT to remember things
V2 (2025): Automatic memory — ChatGPT chose facts to remember from conversations
V3 Dreaming (June 2026): Dynamic, self-updating, multi-source memory synthesis

What changed for users

For Free tier users

Dreaming V3 is now available on the free tier (this was the big announcement in the June 4 rollout)
Core dreaming features: auto-updating memories, summary view, manual edits
Smaller memory budget than Plus/Pro
No app connector support

For Plus and Pro users

Larger memory budgets
Connector access (Gmail, Google Drive, files)
Additional privacy controls (per-domain reference rules)
Priority for new memory features

New controls (June 2026)

Privacy and control were key design goals for Dreaming V3:

Memory dashboard: View a readable summary of everything ChatGPT remembers (not a raw list)
Edit/delete individual facts: Selectively remove or correct memories
Per-domain rules: “Don’t use my health memories during work chats”
Global toggle: Turn memory on/off entirely
Temporary Chat: Never uses memory for the current conversation
Export: Download all memory data as JSON

Technical considerations

Scalability

Dreaming V3 was specifically designed for hundreds of millions of users over multi-year timeframes. The architecture uses:

Distributed background processing for memory synthesis
Efficient storage compression for long-term memory
Low-latency retrieval for conversational context
Fallback to no-memory mode when compute is constrained

Privacy implications

Memories are stored per-user with standard OpenAI data controls
Users can delete their entire memory history
Business/Enterprise accounts have additional data retention policies
Memory data is not used for model training (per OpenAI privacy policy)
Connected app data is processed under existing connector privacy agreements

The bottom line

ChatGPT Dreaming V3 represents a significant step toward genuinely personalized AI. Instead of treating each conversation as isolated and requiring users to repeat context, ChatGPT now maintains a continuously updated understanding of who you are, what you’re working on, and what’s changed since your last chat.

The key tension — personalization vs. privacy — is addressed through granular controls and per-domain rules, but users should review their memory settings regularly to ensure the system’s understanding aligns with their preferences.

For developers building on the OpenAI platform, Dreaming means ChatGPT conversations feel more natural and require less context-rewarming, but also means users will expect similar memory capabilities from custom GPTs and API-integrated applications.

Last verified: June 28, 2026. Sources: OpenAI official Dreaming announcement (openai.com/index/chatgpt-memory-dreaming/), OpenAI release notes, Knightli Dreaming analysis, Let’s Data Science coverage.