OpenAI is beginning to roll out Dreaming V3, a redesigned background memory system for ChatGPT, available to Plus and Pro users in the U.S. starting today. Unlike the original saved-memories feature launched in April 2024 — which required users to explicitly prompt ChatGPT to retain specific information — the Dreaming system runs autonomously in the background, synthesizing memories from chat history without any manual commands. Users can review a plain-language summary of what ChatGPT has learned about them, make corrections, and specify topics they want the model to proactively reference. Free and Go users, as well as users outside the U.S., will gain access over the coming weeks.
The update targets three longstanding pain points with earlier memory systems. First, context continuity: ChatGPT can now recall details from past projects without users re-introducing themselves each session — for instance, referencing a user’s specific camera and housing setup when asked about compatible gear. Second, preference persistence: personal constraints like dietary restrictions or travel priorities are applied consistently across unrelated conversations. Third, temporal freshness: rather than leaving outdated context frozen (such as “the user is in Singapore” weeks after a trip ends), Dreaming automatically revises memories as time passes. OpenAI also said recent engineering work cut the compute cost of serving Dreaming to Free users by approximately 5x, making the broader rollout viable. Dreaming debuted as a supplemental add-on to saved memories in April 2025; V3 now replaces saved memories as the standalone primary memory foundation across all tiers.