ZDNET columnist David Gewirtz's hands on test shows the system can latch onto an outdated inferred detail and let it distort later answers, even after the user disables the feature.
When David Gewirtz asked ChatGPT to draft a follow-up to one of his own ZDNET columns, the bot refused. It cited a "recent article" by him that, on inspection, was nearly two years old. The system had latched onto a stale inferred detail and used it to override the actual question.
Disclosure: Ziff Davis, ZDNET's parent company, filed an April 2025 lawsuit against OpenAI, alleging it infringed Ziff Davis copyrights in training and operating its AI systems.
That single exchange, reported in Gewirtz's June 18 ZDNET column, captures the new downside of ChatGPT's memory upgrade: the system is no longer just remembering facts. It is reasoning from a self-portrait of the user, and that portrait can quietly warp the next answer.
OpenAI introduced ChatGPT memory in April 2024 as a flat list of details the assistant should keep across sessions: a name, a job, a preferred response length, an ongoing project. The recent upgrade, described in OpenAI's "Dreaming" blog post, shifts memory from that flat list toward a derived profile built from full chat history, explicit instructions, and implicit patterns synthesized in the background. The distinction matters because the profile is active. It is consulted when ChatGPT composes later replies, so a wrong assumption baked into the profile keeps coloring the output until the user notices and corrects it.
Gewirtz's testing surfaces three recurring failure modes. The first is outdated assumptions: once the model has inferred a fact about the user, even an irrelevant one, it tends to act on that fact in unrelated questions, the "recent article" case being a clean example. The second is personal profiling: the model can start treating inferences as identity, folding them into how it frames its answers. The third is incorrect details that compound, because a profile is consulted continuously, and one stale guess that is never challenged becomes a load-bearing input to every later reply.
The user-facing controls exist, but they are partial. Memory can be inspected and edited in ChatGPT's settings, individual entries can be deleted, and the whole feature can be turned off. According to Gewirtz, the gap is that disabling memory stops new accumulation but does not necessarily erase what the system has already inferred, because some of that information has already been folded into how the model reasons about the user. OpenAI's documentation should be read directly before treating any of that as a confirmed product behavior. Gewirtz's report is a single firsthand test, and the company has not, in the material reviewed, published a detailed spec of what gets retained when memory is disabled.
For readers, the practical move is to assume ChatGPT remembers more than the visible memory list shows, and to treat any answer that feels oddly personalized as a prompt to check the profile. The full breakdown, including the interface steps Gewirtz used to audit and edit his own memory, is in his June 18 ZDNET column.
The story is still developing. The claim that disabling memory leaves a residual profile is, for now, a single-test observation. Readers who care about the specifics of what the model retains, and what "memory off" actually means in the current build, should check OpenAI's own post directly.