# OpenAI rolls out a more scalable Dreaming memory system for ChatGPT > The new background memory architecture synthesizes changing user context, exposes a reviewable summary, and reduces compute for wider rollout. Canonical URL: https://www.devobs.io/news/news-openai-dreaming-chatgpt-memory/ By: Sofia Reyes Published: 2026-09-06T11:58:54.623Z Updated: 2026-09-06T11:58:54.623Z Event date: 2026-06-04 Section: AI OpenAI began rolling out [a new Dreaming memory architecture for ChatGPT](https://openai.com/index/chatgpt-memory-dreaming/) on June 4. Dreaming synthesizes context from conversations in the background, updating its representation as preferences, projects, and time-sensitive facts change. The release started with Plus and Pro users in the United States, with additional regions and lower-cost plans planned afterward. ## Memory becomes a maintained state Earlier saved memories depended heavily on explicit cues and could grow stale. OpenAI’s first Dreaming version supplemented that list by curating information from chat history. The 2026 architecture makes the synthesized memory more capable and compute-efficient, with a summary page where users can inspect highlights, add or correct details, and give instructions about what should be surfaced. OpenAI evaluates the system along three axes: carrying useful context into later chats, following preferences and constraints, and updating facts as time passes. The last property is particularly important. A trip planned for next month should eventually become a past trip rather than remain a current location signal. OpenAI says serving improvements reduced the compute required for Dreaming enough to support a broader rollout. ## Product teams need temporal tests A memory system is not just retrieval. It performs consolidation, conflict resolution, expiry, and interpretation. Teams building similar features should test contradictory preferences, recurring events, deleted chats, changing names and locations, and long periods without interaction. The visible summary needs clear semantics: whether it is the complete stored state, a generated view, or a subset selected for display. Users also need a reliable way to correct and suppress context. Evaluate whether deletions propagate to derived memories, whether temporary sessions remain isolated, and how personalization behaves when a fact is uncertain. Memory improves continuity only when freshness and user control are part of the same engineering design. ## Source references -