I’ve noticed a substantial drop in ChatGPT’s ability to handle long-running conversations after a recent update.
This is not about factual knowledge, benchmark-style reasoning, or occasional hallucinations. The problem is specifically with continuity of thought across a long conversation.
Until recently, ChatGPT was very good at following an ongoing line of reasoning over many turns. It could remember why a comparison was being made, understand how earlier conclusions affected later questions, and generally behave as if it had actually followed the conversation rather than simply retrieving isolated facts from it.
Recently, that behavior seems noticeably worse.
The most common failure pattern I am seeing is:
The conversation establishes a complex context over many turns.
I make a follow-up comment that depends heavily on that context.
Instead of continuing the existing reasoning, ChatGPT extracts a few recent keywords and generates a new generic interpretation.
When corrected, it apologizes and restates the correction, but often fails to absorb the underlying reason why the previous answer was wrong.
On later turns, the same kind of mistake reappears in a slightly different form.
What makes this particularly frustrating is that the model may still recall many individual facts correctly. It can list project status, dates, previous decisions, or other details, yet still fail to understand the actual conversational context connecting those facts.
In other words, factual recall can appear intact while semantic continuity is broken.
A good analogy would be a student who has copied every item from the class notes into a checklist, but has not actually understood the lesson. If asked to summarize the notes, the student performs well. If asked to continue the argument that was being developed during the lesson, the student suddenly behaves as if they were not present.
Another issue is that corrections seem to be treated as new isolated instructions rather than updates to the model’s understanding of the conversation. For example, if I explain that the reason an earlier answer was wrong was because the model misunderstood the purpose of a comparison, it may correctly repeat that explanation afterward, but then continue to reason in the same shallow, keyword-driven way.
This creates a strange experience where the model can produce an extremely detailed “state summary” of the conversation while still failing at the much simpler task of understanding what I am actually talking about.
The degradation is especially noticeable in long-term creative, development, research, or strategy conversations, where the value of ChatGPT comes from cumulative understanding rather than isolated Q&A.
Previously, the model often felt capable of maintaining a mental model of:
why earlier decisions were made,
what assumptions had already been rejected,
how I tend to evaluate certain situations,
why two things were being compared,
and which parts of the discussion were important versus incidental.
Now it often feels more like:
“retrieve some recent context → classify the current message → generate a plausible response.”
That is a major regression for long-running conversations.
I cannot say whether this is caused by a model update, context-management changes, summarization behavior, routing, or something else. But from a user perspective, the difference is significant and very noticeable.
The biggest issue is not simply that the model forgets things.
It is that it increasingly seems to remember the nouns while forgetting the reasoning.
For users who rely on ChatGPT as a long-term thinking partner, that distinction matters a lot.
Related discussions:
ChatGPT Overvalues Professional-Looking Sources and Amplifies Their Errors
Request to Improve the Quality of GPT Search and Research
ChatGPT Plus conversations reaching maximum length unusually fast across multiple new chats
Severe regression in long-conversation context retention and reasoning quality after a recent ChatGPT update
(Why can’t I include links here?)