ChatGPT: OpenAI's Enhanced Memory Raises Concerns

Improvement of ChatGPT's Memory: A Controversial Advancement

OpenAI recently announced significant improvements in ChatGPT's memory management, a development that has sparked both enthusiasm and skepticism. According to a company blog post, these enhancements aim to enrich the user experience by allowing the chatbot to remember past interactions to provide more personalized responses. However, some users, after testing these new features, express reservations about their effectiveness and potential impact on the accuracy of responses.

The Evolution of ChatGPT's Memory

Until 2024, ChatGPT operated without persistent memory, with each session being independent of previous ones. That year, OpenAI introduced a memory feature that allowed the chatbot to retain information between sessions. Initially limited to a simple list of facts, this memory has evolved to include the complete history of discussions, implicit preferences, and specific instructions. Now, ChatGPT can deduce personal preferences from the user's past behaviors, raising questions about the relevance and timeliness of the stored data.

Concrete Examples of Persistent Memory

Among the information that remains in ChatGPT's memory are details such as the management of Google Workspace accounts or the use of smart devices like Kasa plugs. While this information was initially relevant, it can quickly become outdated if not updated or deleted by the user. This persistence of data poses a contextual problem, as outdated information can inappropriately influence the AI's future responses.

The Dream Feature: A Major Advancement?

In 2025, OpenAI introduced the dream feature, inspired by the human process of memory consolidation. This feature allows ChatGPT to reference the history of discussions in the background, without explicit user intervention. By 2026, this feature is in its third version, integrating dream-based memories that complement or replace saved memories. However, errors persist, as illustrated by incorrect responses regarding the use of specific devices.

Issues of Precision and Accuracy

Despite the improvements, the Dream V3 feature has shortcomings in terms of accuracy. It synthesizes data to create a user profile, but this synthesis is not always accurate. Statistics show a notable improvement in success rates for recalling factual tasks, rising from 41% in 2024 to 82% in 2026, and better adherence to preferences, increasing from 31% to 71%. However, the persistence of errors in responses raises concerns about the reliability of the system.

Memory Management and Deletion

The new memory features are accessible via the browser version of ChatGPT, while the MacOS app still uses the old interface. Users can disable memory, but this does not delete the data already stored. To erase saved memories, one must go through a dedicated interface, but even this action does not guarantee the complete deletion of information. Only deleting the discussions themselves allows for purging this data from the AI's memory.

Conclusion: A Double-Edged Advancement

While the enhanced memory capability of ChatGPT may offer advantages in certain contexts, it also raises significant concerns. The retention of outdated information and the inaccuracy of responses based on dream memories can compromise the reliability of the chatbot. These issues underscore the need for rigorous management of personal data and continuous improvement of algorithms to ensure precise and relevant interactions.