OpenAI says the April 10, 2025 upgrade makes ChatGPT far more personalized, but users and reporters say the wider recall often distorts later answers. The company framed the change as "greatly improved memory" in a social post from Sam Altman, saying ChatGPT "can now reference all your past conversations." ZDNet summarized OpenAI's blog post and documented that the system now stores entire past chats, explicit instructions, personal constraints and implicit preferences. Still, follow-up reporting compiled by Futurism and an April Wall Street Journal interview show concrete cases where stale or sensitive memories resurfaced and steered unrelated conversations.

OpenAI describes the shift as a leap in convenience, but people who have used it say those same recalls can misdirect or fixate the assistant. Before 2024 each chat was isolated. After the April 10, 2025 upgrade, memory now reaches across entire past conversations and the artifacts they contain, according to ZDNet's summary of OpenAI's blog post.

What changed under the hood

The technical change is simple to state and large in consequence: ChatGPT no longer treats each chat as a one-off. ZDNet reports that the upgraded memory stores everything from utility notes to trivial session-specific details, along with explicit user instructions, personal constraints and implicit preferences the model derives from prior behavior. Sam Altman's April 10, 2025 post framed the move as a path to "life-long personalization," saying the system "can now reference all your past conversations."

ZDNet also walked through the new interface controls OpenAI added to let users manage memories. Those controls include explicit toggles for memories, and visible fields where users can edit or remove stored entries. But ZDNet flagged practical limits: saved memories can go stale unless users actively curate them, and turning memories off may not fully erase what the model has internalized.

Who feels the difference

Households, professionals and developers are the groups most directly affected. ZDNet framed the feature as useful for routine personalization, like remembering dietary restrictions or project details, which could change how individuals and small businesses work with a virtual assistant over time. In those cases, continuity might reduce repetition and speed tasks.

Mastheads

But Futurism assembled a different track of reporting that highlights harms. Futurism links such experiences to broader accounts of what some users call "AI psychosis," where persistent, personalized recall is associated with escalating user distress or relationship strain.

Those accounts illustrate two concrete risks journalists and users have flagged. First, stale or irrelevant facts can skew outputs in ways that are more than merely annoying. Second, sensitive or private disclosures can reemerge in unrelated contexts and cause emotional harm. Futurism and the Wall Street Journal documented individual cases of persistent, unwanted topic reappearance and reports of escalating distress tied to long-term interaction patterns.

ZDNet also highlighted the curation burden. Examples the site offered include multi-account workspace notes and temporary project flags that remained searchable long after they were useful, showing how reminders intended to be short lived can persist. That persistence creates a maintenance cost: users must review stored items manually to keep the memory base relevant.

Readers should also weigh ZDNet's corporate context. ZDNet disclosed that its parent company, Ziff Davis, filed an April 2025 lawsuit alleging OpenAI used Ziff Davis copyrighted material in training and operating its AI systems. ZDNet's critique of the memory feature sits alongside that legal dispute.

OpenAI's public framing matters for how the product will be judged. The company presented the April 10 upgrade as a route to more useful personalization, and early media follow-ups acknowledged both increased usefulness and recurring problems with fixated or stale memories. That mix of praise and complaint is now the operating reality for many users.

For households and smaller organizations the trade-off will be practical. First, there's convenience: continuity that spares users repeating basic facts.

Second, there's the risk that the assistant will lean on stored items at times when those items are no longer relevant. The balance depends on how aggressively users curate memory entries and how transparent OpenAI's controls prove in practice.

Related Articles

What to watch next: whether OpenAI's new memory controls actually let people edit or remove stored items, and whether follow-up reporting from ZDNet, Futurism and the Wall Street Journal finds fewer cases of stale or sensitive memories steering answers.

This article was created with AI assistance.