OpenAI recently rolled out a significant “upgrade” to ChatGPT’s memory capabilities, promising a more personalized and seamless user experience. However, my hands-on testing suggests these improvements might actually be a double-edged sword. While intended to enhance relevance, they introduce a host of issues from outdated assumptions to outright incorrect profiling, subtly — and sometimes overtly — distorting the AI’s future responses.
This isn’t just about the chatbot remembering a few facts; it’s a comprehensive system that now tracks your entire chat history, explicit instructions, personal constraints, and even implicit preferences derived from casual remarks. The implications are profound, raising concerns about the accuracy and reliability of information we receive. Could this “smarter” memory actually make ChatGPT less trustworthy?
The Evolution of ChatGPT’s Memory
Before 2024, ChatGPT operated without any persistent memory; each chat session was a standalone interaction. Whatever information you shared remained confined to that single conversation, unavailable to any subsequent exchanges. This ensured a clean slate, but also meant you had to repeatedly provide context.
The first iteration of “memories” arrived in 2024, essentially a list of explicit facts you could save. While a step towards personalization, these early memories quickly became stale. They lacked implicit context and often remained relevant only for the specific session they were created in, without a mechanism to update or discard outdated information.
Then came “Dreaming” in 2025, a feature designed to allow the model to reference your chat history in the background without explicit prompts. This marked a significant shift, as ChatGPT began to automatically curate memories, combining saved facts with insights from past conversations. This initial version of Dreaming (V0) laid the groundwork for a more autonomously evolving memory system.
Fast forward to 2026, and we now have Dreaming V3, which has largely replaced or augmented the older saved memories. This latest iteration doesn’t just scan history; it performs data synthesis, effectively compiling a detailed, multi-layered “dossier” about you. OpenAI claims this version can carry forward complex context and track long-running projects with remarkable efficiency.
- Factual task recall success jumped from 41% in 2024 to 82% in 2026.
- Ability to stay correct over time improved from a 9% task rate in 2024 to 75% in 2026.
- Preference adherence rose from 31% in 2024 to 71% in 2026.
Despite these impressive metrics, my personal experience reveals concerning inaccuracies. When asked about my “experience with Kasa,” ChatGPT stated I used specific smart plugs for energy monitoring and later integrated them into Home Assistant. The Kasa plug detail was vaguely correct from a past discussion, but the Home Assistant integration was entirely false—I’ve never even installed it.
OpenAI acknowledged this, explaining that what users see is a “high-level memory summary” designed for review, not a complete inventory. While it might not show every detail, the underlying context can still influence conversations. This suggests a powerful, yet opaque, system at play.
Managing Your AI Memories
The new memory features are primarily accessible through the browser version of ChatGPT for Plus and Pro tier subscribers, with a rollout planned for all users soon. To find them, navigate to Settings > Personalization > Memory section. Here, you’ll find options to manage how ChatGPT remembers you.
You can partially disable the memory capability, preventing new dream-based consolidation, but this won’t remove already stored memories or your chat history. To delete specific memories, you must go into the saved memories interface and purge them manually. Even then, complete data deletion often requires deleting the entire chat session itself.
An important “gotcha” to note: turning off Memory/Personalization does not disable safety features. OpenAI states that in “rare, high-risk situations,” ChatGPT may still use limited, safety-relevant context to respond more safely. While this isn’t reported to anyone, it means certain sensitive information might persist, adding another layer of complexity to data management.
Beyond deletion, the “Manage” button under Personalization allows you to refine the consolidated profile ChatGPT has built about you. This interface, often presented as a narrative of interests and preferences, lets you select aspects and mark them as “Don’t mention this again” or add comments for clarification. However, this relies on ChatGPT’s often-flawed interpretation of your persona.
Why This Memory Upgrade Worries Me
While the technical prowess behind Dreaming V3 is undeniable—it boasts a 5X reduction in compute cost, making it scalable for mass access—its implementation raises serious concerns. The AI is constantly revising its internal “mental state,” tracking timestamps and theoretically experiencing the passage of time alongside you. Yet, my tests show it struggles to keep up with actual life changes.
The feature retains wildly out-of-date information and uses it to filter current responses. ChatGPT’s “dossier” about me, derived from past conversations, is often inaccurate, much like a friend who still pegs you to an outdated phase of your life. This isn’t just a minor annoyance; it’s a fundamental flaw.
Furthermore, not all my conversations are about me. I often use ChatGPT for research, and these inquiries might be misinterpreted as personal preferences, attaching irrelevant data to my profile. This means the AI never truly knows who I am; it only knows what it *thinks* it knows, based on potentially misconstrued interactions.
This increases our cognitive burden, forcing us to constantly verify the veracity and completeness of every AI response against its potentially warped view of us. Will ChatGPT omit crucial information because it believes we aren’t interested? Will it modify its presentation because it assumes a certain format preference? This “memory upgrade,” while a technical triumph, borders on irresponsible, introducing biases and inaccuracies that are difficult, if not impossible, to prune.
Source: ZDNet – AI
