It has limited ‘permanent memory’ it can ‘update’ for ‘special’ user requests. eg. for months I asked chatGPT to stop recommending therapy – to no avail

Understanding the Limitations of AI Memory and User Interaction: A Case Study with ChatGPT

In the evolving landscape of artificial intelligence, understanding how models like ChatGPT handle memory and user requests is crucial. While AI language models have become increasingly sophisticated, they still operate within certain constraints—particularly regarding how they remember and incorporate information over time.

A personal anecdote sheds light on these limitations. Over several months, a user attempted to persuade ChatGPT to cease recommending therapy, a request that went unfulfilled repeatedly. This highlights that, despite ongoing interactions, ChatGPT does not possess truly permanent or updateable memory related to individual user requests.

However, an interesting turning point occurred when the user introduced definitive personal information, including a fabricated letter from their psychiatrist explicitly stating “no therapy.” By asserting this authoritative document, the user effectively demonstrated how external, authoritative inputs might influence the AI’s responses within its operational boundaries. ChatGPT responded by indicating an action—such as adding this information to its “permanent memory”—which, while not genuinely altering the model’s core memory, suggests its ability to incorporate context temporarily or under certain conditions.

It’s important to clarify that ChatGPT does not have actual long-term memory of past interactions unless designed explicitly with such a feature—something current models do not support by default for privacy and design reasons. Instead, they process each session or prompt independently, with temporary context retention during a conversation.

The user also shares a humorous note about restrictions surrounding the generation of certain images—specifically, content related to deities—highlighting AI models’ content policies and how certain groups or inputs receive special consideration due to legal or policy reasons. This underscores ongoing challenges in balancing creativity, legality, and ethical considerations in AI-generated content.

Interestingly, the user references a broader perspective—that sometimes, achieving desired outcomes might involve “kicking off” or deliberately initiating certain actions or interactions, akin to forking a repository in open-source development. This analogy illustrates their view that the AI landscape is like a community-driven project, where forks and collaborative efforts represent avenues for modification and improvement.

In summary, this case illustrates the current state of AI models like ChatGPT—powerful but bounded by design in terms of memory and user request handling. While users can influence responses temporarily and use external authoritative information to guide interactions, true long-term memory without explicit system design remains an ongoing challenge. The evolving nature of AI development continues to open new possibilities while reminding us of existing limitations.

For developers and users alike, understanding these constraints is essential for effective and ethical engagement with AI technologies. As the community strives toward more advanced and memory-capable models, discussions around transparency, control, and policy will remain central.

Author’s note: The personal accounts referenced are illustrative and part of ongoing explorations into AI capabilities. For more insights, visit my GitHub profile under the username lumixdeee.