OpenAI launches Dreaming V3 memory system for ChatGPT on June 4
In brief
- Dreaming V3 memory architecture launched June 4 for ChatGPT Plus and Pro subscribers in the US
- System improves context retention and conversational continuity across conversations separated by days or weeks
- Three core dimensions: information freshness, thread continuity, and irrelevant noise filtering
- Users control memory viewing, editing, deletion; temporary chats opt out entirely
- Expansion to additional tiers and global regions planned
Memory Evolution
ChatGPT's memory journey began in February 2024 with a basic "saved memories" feature, where users could explicitly tell the model what to remember. In April 2025, the model gained the ability to reference all previous chat history, not just the memories users explicitly saved. Dreaming V3 represents the next step forward in this progression.
The new architecture lets ChatGPT maintain fresher, more continuous context across conversations. The system synthesizes context more effectively across those conversations, understanding how mentioned items fit into the broader arc of ongoing work. Rather than simply recalling that you discussed something, Dreaming V3 grasps how it connects to your larger projects and priorities.
Three Core Optimizations
The optimizations focus on three core dimensions: freshness, continuity, and relevance. Freshness means the system prioritizes recent and up-to-date information over outdated context. Continuity ensures threads of conversation feel connected even when separated by days or weeks. Relevance filters out noise, surfacing only the context that actually matters for the current exchange.
This layered approach keeps the model from drowning in historical data while still maintaining a coherent sense of your work and preferences.
User Control and Opt-Out
OpenAI is giving users the ability to view, edit, or delete their stored memories. For users who want to interact without building up a memory profile, temporary chats offer a full opt-out. Nothing from those conversations gets stored or referenced in future exchanges.
Free-tier access is also reportedly in progress, meaning the memory system won't remain locked behind a paywall indefinitely. The rollout strategy mirrors OpenAI's typical approach: test with paying users first, refine based on feedback, then broaden access.
