OpenAI rolls out 'Dreaming V3,' a self-updating ChatGPT memory engine

Dreaming V3 marks a shift from explicit, user-managed memory to autonomous, model-managed memory. Instead of a list of facts the user manually saves, ChatGPT now continuously distills long-term context from prior conversations and rewrites those memories as circumstances change. OpenAI's internal benchmarks claim factual recall of 82.8% and preference adherence of 71.3%, both meaningful jumps over the prior system.

The self-updating mechanic is the headline feature: a memory like 'you're planning a trip to Singapore in July' automatically becomes 'you went to Singapore in July 2026' once the date passes — keeping the model's internal user model temporally consistent without manual cleanup.

Competitively, this lands directly against Google's Gemini personalization push and Anthropic's projects/memory features, and it deepens lock-in: the more context ChatGPT accumulates, the higher the switching cost. The rollout is gated to Plus and Pro subscribers in the US first.

The caveat that drew attention is the audit trail. By moving from an explicit saved-memories list to an autonomous synthesis layer, OpenAI reduces the transparency users have into exactly what the system 'knows' about them — coverage flagged limits on the audit trail as a privacy concern. Watch how OpenAI handles deletion, export, and correction of auto-generated memories, and whether EU regulators scrutinize the autonomous-profiling angle.