← Back to feed

OpenAI shipped a new version of ChatGPT’s memory system on June 4, 2026, built around what it calls dreaming — a background synthesis process that reviews conversation history and continuously updates what the model knows about you, without waiting for explicit save instructions.

The original saved-memories approach, launched in April 2024, required strong conversational cues to trigger: “remember I’m travelling to Singapore in July” would get stored; most everything else would not. The result was incomplete recall — the system captured explicit requests but missed the implied context that accumulates across hundreds of conversations. Memories also went stale. If circumstances changed, old entries stayed in the list until manually deleted.

The first dreaming update in April 2025 addressed the passive-recall problem by letting ChatGPT reference chat history beyond the saved memories list. The June 2026 update goes further: it runs a dedicated background process that synthesizes memory state across many conversations simultaneously, not just the current one. The system is designed to handle hundreds of millions of users and multi-year time horizons — the scale at which the earlier approach’s correctness and freshness guarantees broke down.

What Changed

Three stated improvements:

• Staleness: Memories are synthesized and updated on an ongoing basis rather than written once and left. Context that no longer applies gets revised rather than accumulating.
• Correctness: The synthesis process cross-references multiple conversations to resolve contradictions before committing facts to memory.
• Coverage: Because the process runs asynchronously over full conversation history, it captures context that emerged naturally without ever being flagged for saving.

The rollout is Plus and Pro users in the US today, expanding to additional countries and Free and Go users over the coming weeks.

The Broader Pattern

The timing is notable. Anthropic shipped its own dreaming mechanism for Claude agents earlier this year, using background replay sessions to let agents learn from prior mistakes between runs. OpenAI’s implementation targets a different problem — persistent user context in a consumer product rather than agent skill transfer — but both labs are converging on the same architectural idea: valuable synthesis happens asynchronously, outside the hot path of a live conversation.

For ChatGPT, this matters most at the long tail of the user relationship. After a year of use, a model that remembers you accurately is meaningfully more useful than one that doesn’t. The question is whether the synthesis quality holds up at the scale ChatGPT operates at. OpenAI has not published latency or recall metrics for the new system.