OpenAI just shipped Dreaming V3, a new memory system for ChatGPT that quietly rewrites how personalization works. Instead of waiting for you to say “remember this,” it runs a background process that learns from your past conversations and synthesizes a living memory state. The name fits. Much like a brain consolidating the day during sleep, Dreaming V3 works through your history during idle time and keeps the useful parts current. It began rolling out to Plus and Pro users in the US on June 4, and recent optimizations cut the compute cost by roughly 5x, which is why free tier access is coming next.
Why this matters for builders
For anyone building on top of LLMs, memory has always been the hard part. Most of us bolt on a vector database, write retrieval logic, and still fight stale or contradictory context. Dreaming V3 pushes that work into the model layer, combining recurrent memory modules with vector recall, and it refreshes what it knows as time passes instead of letting old facts rot.
The catch
Automatic memory means less control. When the model decides what to keep, you need real visibility and deletion controls before you trust it with anything sensitive. Test Dreaming V3 against your own memory stack this week, and if you build agents, watch this closely, because memory as a platform feature changes what you have to build yourself.