← Back to feed

When OpenAI shipped Dreaming 2.0 on June 4, the public framing was a fix for staleness: background synthesis of memories across conversations, cross-referencing to resolve contradictions, asynchronous updates that capture context users never explicitly flagged. The problem the company named was incomplete recall.

Long-term Pro subscribers are reporting a different problem: the explicit control layer is gone.

The original saved-memories system functioned as a notebook. Users could see every stored entry, add items directly, edit inaccuracies, and delete anything outdated. What was in the notebook determined what the model knew about you. The relationship was legible and user-controlled.

Dreaming V3 removes the notebook. The system now decides what is important on your behalf. Pro users with over a year of daily usage report the transition has introduced consistent failures: memories they consider critical require repeated reminders before the system retains them; information that was accurate months ago resurfaces in contexts where it no longer applies; there is no way to inspect or correct what the model currently believes about you.

Why the Architecture Change Causes Regression

The original system’s legibility was a feature, not a limitation. A user who told ChatGPT their surgery date had it stored, visible, and persisted until deleted. If the information became outdated, the user deleted it. The contract was explicit.

An AI-managed synthesis system does not offer that contract. The system’s decisions about what constitutes a significant memory are made asynchronously, across full conversation history, with no user-visible reasoning. A memory that the system deweights — because it hasn’t appeared in recent conversations, because other signals overrode it, because the synthesis process made an incorrect inference — degrades silently.

The failure mode is the inverse of staleness. Staleness means old information persists when it should not. The Dreaming V3 reports describe important information that does not persist at all, while genuinely outdated context returns unprompted. The synthesis process is doing active curation and getting it wrong for some long-tail users.

The Power User Cohort Problem

OpenAI’s stated goal for the system is scale — handling hundreds of millions of users over multi-year time horizons. The synthesis approach makes engineering sense at that scale. A background process that interprets and compresses full conversation history is more tractable than a per-user notebook requiring explicit management.

The users most affected by the tradeoff are the ones with the most history. A long-term Pro subscriber with hundreds of conversations has a rich memory graph that the synthesis system is now responsible for compressing and prioritising. The more context there is, the more opportunities for the synthesis to make incorrect priority judgments.

For a new user, Dreaming V3 likely improves over the baseline — automatic synthesis captures context the notebook would have missed entirely. For a user with 18 months of accumulated history, the regression is measurable: a system that used to reflect stated preferences now reflects an AI inference about what those preferences probably are.

The Broader Pattern

OpenAI is not alone in this tradeoff. Anthropic shipped a dreaming mechanism for Claude agents earlier this year — background replay sessions that let agents learn from prior mistakes between runs. The architectural logic is the same: valuable synthesis happens asynchronously, outside the hot path of a live session.

But ChatGPT’s user base includes a large cohort of heavy personal-use subscribers who built genuine workflows around the explicit notebook. Anthropic’s dreaming is targeted at agents, not at personal memory continuity. OpenAI is trying to use the same architectural approach for a different problem — persistent user identity across consumer use — and the failure surface is different.

The specific complaint from heavy users is that they lost control over ground truth. Under the old system, if you believed the model had the wrong picture of your situation, you could look at the notebook and fix it. Under Dreaming V3, there is no accessible notebook. The model’s picture of you is synthesised from conversation history you cannot directly query, by a process whose reasoning is not surfaced.

OpenAI has not published recall accuracy or freshness metrics for the new system, and has not announced plans to restore manual memory management.