Dreaming: Better memory for a more helpful ChatGPT

OpenAI is rolling out a revamped memory system for ChatGPT, and the pitch is simple: the assistant should stop forgetting who you are every time you open a new tab. The company frames it as 'dreaming,' a nod to how the model consolidates scattered details from past conversations into something coherent it can actually use later, rather than just stuffing a transcript into a context window and hoping for relevance.

The old memory approach leaned heavily on saving explicit facts you told it to remember, plus some quiet inference from chat history. That worked for simple stuff, but it got messy fast. Preferences drift, context changes, and a system that only accumulates facts without reorganizing them eventually becomes a junk drawer of outdated or contradictory notes. OpenAI says this update focuses on keeping what's relevant fresh and letting older, less useful details fade, rather than piling everything up indefinitely.

What's notable here isn't the existence of memory itself, ChatGPT has had some version of this for a while, it's the shift toward something closer to active curation. Instead of a flat list of saved facts, the system is meant to synthesize and update understanding over time, more like a running model of you than a sticky-note pile. That's a meaningfully different engineering problem, and it's one every major assistant maker is quietly wrestling with, because raw context windows don't scale to months of usage.

OpenAI isn't publishing hard numbers on accuracy gains or how much data gets retained versus dropped, and there's no detail yet on user controls for editing or auditing what the model has inferred about you. Given how much scrutiny personalization features get once they touch sensitive personal details, that transparency gap is worth watching as this rolls out more broadly.