ChatGPT's nightly memory consolidation keeps user preferences fresh across sessions

OpenAI rolled out a memory overhaul for ChatGPT this week that uses background consolidation passes to keep user preferences current across sessions. The system, which the company calls "dreaming," runs nightly batch processes on stored context—pruning stale details, reinforcing patterns that recur across conversations, and surfacing preferences that might otherwise get buried under new input. The result is a chatbot that remembers your coding style or dietary restrictions without requiring you to re-state them every time you open a new thread.

The mechanics rely on a nightly batch process that replays recent exchanges and updates a persistent memory store. OpenAI says the approach reduces context drift: instead of treating every conversation as a fresh slate or relying on a static profile card, the system adjusts its internal representation of you as your needs shift. A user who switches from Python to Rust, or from vegetarian to vegan, will see that change propagate forward without manual edits to a settings page.

The feature is live now for Plus and Team subscribers, with Enterprise and Edu tiers following later this month. Free-tier users get read-only access to their memory store but can't trigger the dreaming cycle. OpenAI has not disclosed token costs for the background passes or whether the consolidation runs on the same GPT-4 infrastructure that powers chat itself.

How the system handles conflicting signals remains an open question. If you ask for concise answers one day and verbose explanations the next, it's unclear which preference wins or how long the dreaming cycle takes to notice the shift. The next update should clarify retention windows and give users a way to audit or roll back memory changes that drift off course.