OpenAI Overhauls ChatGPT Memory With 'Dreaming V3'
On June 4, 2026, OpenAI announced it has begun rolling out a more capable memory system in ChatGPT, designed to carry context across conversations and keep it useful over time. On its official announcement page, the company frames the update as "Dreaming V3," saying improvements to memory synthesis optimize for freshness, continuity, and relevance. The rollout starts immediately for Plus and Pro users in the United States, with expansion to additional countries and the Free/Go plans expected within a few weeks (OpenAI).
The system's defining feature is that a background process automatically synthesizes and updates memories from chat history without requiring users to explicitly say "remember." Users can review and edit a summary of their memories and add instructions on a Memory summary page, and an option to switch back to Legacy saved memories is also provided. Plus and Pro users receive "2x more memory," and the feature becomes available on iOS and Android by updating to the latest app version (OpenAI on X).
The memory feature has evolved in stages over the past two years. Saved memories, which required users to explicitly instruct what to remember, debuted in April 2024, and Dreaming V0, which referenced chat history to remember automatically, was introduced in April 2025 (OpenAI). Dreaming V3 extends this lineage, using an independent architecture to sharply improve computational efficiency and to address longstanding challenges such as staleness, correctness, and scalability (OpenAI). OpenAI positions it as a core feature for evolving ChatGPT into a long-term partner that "knows you."
On specifications, memory is intended for high-level preferences and details rather than storing large amounts of verbatim text. In Temporary Chat, referencing and creating new memories can be disabled. Memory sources vary by plan and can include Past chats, Saved memories, Custom instructions, Files, and Gmail (excluding EEA/UK/Switzerland); the feature is offered first on Web, with mobile support expected soon (Memory FAQ). Past memory updates were delayed in regions such as the EU and UK for regulatory reasons, and this time too the U.S. takes the lead (TechCrunch). No specific benchmark figures have been officially released, but improvements are suggested across three axes: carrying context, following preferences, and handling the passage of time (OpenAI).
Reactions on X leaned positive, with many calling it a "big update." Users praised the automatic synthesis, time adaptation, and efficiency gains, with some citing roughly a 5x computational efficiency improvement and arguing it could become a "moat" for multi-year personalization. Competitors such as Claude's similar memory features and Gemini's Spark were cited, with observers noting that "who remembers you best" is becoming a differentiator in consumer AI (TeksCreate on X). Others were skeptical, arguing this is not true continuous learning (weight updates) and questioning whether the Legacy switch option signals a lack of confidence in the new system (Redwinelips on X), while practical use cases such as extracting memories from natural conversation and updating state after trip planning were offered as examples (coder_left on X). The balance between expectations for full automation and users' desire for control is at the heart of the debate.