I would like to submit feedback regarding the changes to ChatGPT’s memory functionality introduced in June 2026.
In my experience, the quality and reliability of ChatGPT’s work have deteriorated materially since this rollout, particularly in tasks that depend on persistent, precise working instructions. The issue is not primarily whether ChatGPT can remember biographical facts or general preferences. The critical loss is the user’s ability to establish and maintain granular, durable operating instructions that govern how substantive work is performed.
Before the change, I could use memory to establish persistent requirements concerning matters such as language, drafting style, evidentiary standards, reasoning discipline, formatting, and the treatment of formal or process-sensitive work. Since the rollout, ChatGPT has repeatedly failed to follow instructions of exactly this kind. Examples include disregarding explicit linguistic requirements, departing from specified standards of reasoning and evidence, introducing unsupported propositions, and providing isolated drafting suggestions despite explicit instructions that amendments must be integrated into the complete document.
The operational consequence is significant. Work that previously required relatively few iterations now requires repeated correction simply to restore instructions that had already been established. This creates a substantial efficiency loss for the user. It also appears economically inefficient for OpenAI: every unnecessary corrective iteration requires additional model inference and therefore additional computation. A product change that causes users to spend more time and OpenAI to spend more compute to obtain the same output quality is a functional regression, not an improvement.
The current alternatives do not adequately solve the problem. A single general-purpose Custom Instructions field is too coarse for users who work across materially different contexts. Nor is an opaque or largely model-managed memory system an adequate substitute for explicit, user-controlled persistent instructions. For professional, academic, analytical, legal, financial, or other high-precision workflows, users need to be able to define durable operating constraints themselves and to know that those constraints remain in force unless they deliberately amend or remove them.
This is also a Responsible AI issue. Responsible AI is not limited to content safety. It also requires predictability, controllability, transparency, reproducibility, and reliable adherence to user-defined constraints. Where a user is preparing formal or consequential work, repeated non-compliance with previously established instructions can introduce factual, procedural, evidentiary, or drafting errors. Removing effective user control over persistent instructions therefore increases rather than reduces operational risk.
I therefore strongly recommend that OpenAI either reverse this aspect of the June 2026 rollout or restore, in a subsequent release, a first-class mechanism for explicit user-managed persistent memories and operating instructions. Users should be able to create, edit, categorise, enable, disable and delete individual persistent instructions directly, rather than relying predominantly on model-managed memory or a single undifferentiated Custom Instructions field.
The underlying models continue to improve substantially. It is therefore particularly difficult to justify a surrounding product-layer change that makes the core model less usable, less predictable and more expensive to operate in real workflows. The memory redesign should not negate improvements in the underlying model by increasing the number of iterations required to obtain compliant output.
This should be treated as a product regression and addressed accordingly.