ChatGPT's dreaming memory system rolls out

OpenAI has been developing memory features for years. Memory first launched in April 2024, allowing users to save explicit notes. An upgrade in April 2025 introduced the first version of dreaming, enabling automatic memory curation. Until now, dreaming supplemented saved memories but was not sufficient alone. The new system, dreaming V3, is more powerful and compute-efficient, built on the same background process.

The improvement is measurable across three objectives: carrying forward context, following preferences, and staying current over time. The article shows examples where the new system recalls camera setups, applies user preferences like wildlife photography, and updates events that have passed. Evaluations indicate the dreaming-based system improves the model's ability to recall relevant facts, but specific scores are not disclosed.

The rollout is phased, starting with Plus and Pro users in the US. Over coming weeks, it will expand to other countries and Free and Go users. This gradual approach may help OpenAI manage the scalability challenges noted with hundreds of millions of users. However, the true test is whether the system remains useful and correct over multi-year usage. If dreaming meets expectations, memory could become a stronger differentiator for ChatGPT compared to competitors without similar robust memory features.