Follow ZDNET: Add us as a preferred source on Google.

ZDNET's key takeaways

ChatGPT's memory now builds a profile from past chats.
Old or irrelevant details can distort future AI answers.
Turning memory off may not fully erase what ChatGPT knows.

OpenAI recently shared details on enhancements to how ChatGPT handles user memories, and the updates appear substantial. Yet after reviewing them closely, I find the changes concerning. It remains unclear if disabling the new options improves or worsens the situation.

Memories in this context refer to information shared during conversations with ChatGPT. Launched in 2024 as a simple collection of facts, the system has grown to encompass full chat histories, custom instructions, personal limits, and inferred preferences drawn from interactions and offhand comments.

This piece covers three main areas. It begins with technical details from OpenAI's latest announcement on memory upgrades. Next comes an overview of the controls available to adjust these memories. It concludes with an explanation of why the feature raises red flags for me and potentially for others.

Let's examine the details.

Processing Background Context

Prior to 2024, ChatGPT operated without any form of memory. Individual sessions remained isolated, with no carryover of details to future interactions.

In 2024, the initial memory system arrived as a basic list of facts. This option persists today, though much of the stored content proves outdated or irrelevant. Three examples from my own list include references to managing multiple workspace accounts for storage needs, tracking positions in a coding project involving repeated elements, and setting up a particular smart home device.

These entries linger despite applying only to single past sessions from years earlier. The list lacks awareness of context or timing, allowing facts to become obsolete without automatic updates unless manually reviewed.

By 2025, OpenAI introduced background processing referred to as dreaming. This allowed the model to draw from chat history automatically rather than requiring direct prompts. It began generating memories on its own, combining the original saved facts with this new layer.

Now in 2026, the dreaming capability reaches version 3. Saved memories continue but are often supplemented or supplanted by synthesized versions from the dreaming process.

For instance, a query about experience with a smart plug brand prompted a response citing a specific model for energy tracking and a later shift to another platform. Neither detail matched actual history, as no such platform had ever been used and the device remained unused in storage.

OpenAI clarified that the output represents a high-level summary rather than an exhaustive list, intended to simplify review while omitting some technical specifics that might still influence replies.

The current version synthesizes data into an overall profile that may contain errors. The company reports gains in accurate recall, long-term consistency, and alignment with stated preferences across versions.

Compute requirements for this ongoing analysis dropped significantly, enabling broader availability. The system now updates its internal view continuously, incorporating time-based markers from conversations.

Access to version 3 is live for paid subscribers, with expansion to free accounts planned soon.

Exploring the Controls

The updated memory options appear in the browser interface under Settings and Personalization. The desktop application retains the prior layout for now.

Memory can be toggled off, yet existing stored items and chat history remain intact. Background processing stops, but prior data persists.

Full removal requires deleting items from the saved list individually. Chat sessions themselves must also be cleared to eliminate their influence entirely.

A note in the documentation indicates that safety-related elements may still draw on limited context even when memory features are disabled, to support safer responses in specific cases.

An additional manage section allows edits to the compiled profile of interests and habits. Items can be flagged to avoid future mentions or annotated with corrections.

Reflecting on the Implications

Effective memory features prove valuable in certain workflows, such as maintaining context across coding tasks or server-based projects. Custom instructions help retain details between sessions in those setups.

Nevertheless, the approach in ChatGPT leaves me uneasy. Outdated entries continue to shape outputs, and even the advanced version can introduce false details, such as assuming use of a home automation system that was never installed.

The system applies a personal filter derived from limited inputs, leading to assumptions that may not reflect reality. This resembles relying on incomplete snapshots like social posts or outdated recollections from others.

Research queries unrelated to personal life risk being folded into the profile, altering how future answers are framed. The full scope of inferred knowledge stays hidden, making verification difficult.

While the technical progress in scale and efficiency stands out, the feature carries risks. It draws from older sessions created under different expectations about data use. Pruning remains cumbersome, and updates may lag behind actual changes in user circumstances.

Users now face added effort to cross-check responses against their full interaction history. Not everyone will apply the same scrutiny, potentially allowing skewed or incomplete information to go unnoticed.