OpenAI is updating ChatGPT’s memory feature, making it more reliable and extending access to free users for the first time. The system, which lets the chatbot retain details about individual preferences and conversation history, has long struggled with maintaining accuracy over extended periods and across millions of users. This latest refinement attempts to address those practical limitations.
The core improvement centers on an enhanced “dreaming” process, OpenAI’s term for how the model synthesizes and organizes stored information. Previously, memory could grow stale or inconsistent, especially as interactions stretched over months or years. The new architecture aims for better scalability while using less computing power. Users can now review a summarized overview of what the model has retained, edit entries, or add specific instructions about topics to reference or avoid. This offers a more transparent way to manage the personal data the system holds.
Three main areas show noticeable progress. The updated memory better carries forward relevant context from past exchanges, respects stated preferences and constraints more consistently, and adjusts for the passage of time—preventing outdated details from lingering awkwardly in responses. These changes reflect ongoing efforts to make conversational AI feel less like a reset machine and more like a persistent assistant. Yet challenges remain; no system is immune to occasional inaccuracies or the risk of over-relying on stored assumptions that users might later regret sharing.
For paying Plus and Pro subscribers in the United States, the upgraded memory system begins rolling out immediately, with storage capacity doubled to accommodate more retained information. Free users will receive a version of this dreaming-based memory over the coming weeks, following significant efficiency gains—reportedly around a fivefold reduction in required computing resources. This expansion brings free accounts closer to the persistent context that subscribers have tested for some time, though likely with certain limits to maintain service stability at scale.
The move fits a broader pattern in AI development. Early chatbots relied on short-term conversation history that vanished between sessions. Memory features represent a step toward more natural, ongoing interactions, but they also raise familiar questions about data privacy, consent, and the long-term consequences of machines building detailed user profiles. OpenAI’s emphasis on reviewable summaries is a welcome acknowledgment of these issues, though users would be wise to treat any AI-held personal information with healthy skepticism.
As ChatGPT’s memory feature evolves, it highlights both the potential and the constraints of current large language models. Improved retention can make everyday tasks smoother, from project continuity to personalized recommendations, yet it also underscores how much work remains before these systems handle nuance and reliability at the level many users expect. The gradual rollout to free users suggests OpenAI is balancing ambition with operational reality, a pragmatic approach in an industry often criticized for overpromising.
