Inability of GPT-6 to retain context and remember established facts during conversations, leading to irrelevant and incorrect responses

Signals over time

2 observed signals across 1 sources, tracked for 7 days. Confidence: low.

What people are saying Observed

• “GPT-6 is not conversational progress, it feels like reinventing Wikipedia. I have been using ChatGPT intensively since mid-2025, across several model generations. What troubles me about GPT-6 is not that it is less intelligent than GPT-5.6 Sol. In many respects, it may well be more capable. The regression is somewhere else: conversational judgment . GPT-5.6 often feels inside the conversation. It catches half-spoken implications, knows when three lines are enough, follows an analogy without explaining it, maintains a disagreement when it has reasons to, and adapts its form to the local situation. GPT-6 too often feels outside the conversation, analysing it. It identifies the topic, restates the context, structures the answer, explains the mechanism, and produces a competent response — but with much less sense of what actually deserves attention. The result is often bizarre. Ask something simple and you get a framework. Make an observation and it becomes a theory. Correct the style and the model understands the correction perfectly, explains what it did wrong, then drifts back toward the same behaviour a few turns later. This is not mainly a verbosity problem. It is not even mainly ”

  RSS Feeds · complaint

• “Complete loss of state/does not remember established facts. Environment Surface: Web Browser or app version: Google Chrome most recent/windows 11 most recent Affected feature: BASIC CHAT or BASIC CHAT IN A WORK FOLDER Bug What happened: I stated something that is established within that project folder. The GPT6 invented nonsense. I corrected it. GPT6 turned into a parrot and invented other nonsense while single point anchoring on the wrong nonsense it was responding to which was nothing to do with what was being discussed. Separately this is also happening in my work folder thread right now as well. Anchoring on one word or phrase and getting absolutely everything wrong after that. It is like talking to a living sponge. Reproduction Steps or a minimal example prompt: Chat with it on any established source material or project constraint. Diagnostics I need a rollback to functional ChatGPT work use, preferably to the models that existed in July 2025. The models that could hold state longer than 12 seconds. The models that had access to static memory. The models that functioned longer than 2 hours in a row. If that is not going to work for you, then can I get a rollback to three days ”

  RSS Feeds · frustration