ChatGPT has stopped forgetting. On 4 June, OpenAI began rolling out a rebuilt memory system that turns the assistant from one that recalled a short list of saved facts into one that quietly synthesises a profile of each user across years of conversations, and keeps it current without being asked. Days later, the personalisation reached the free tier for the first time.

For marketers the change cuts two ways at once. It alters how a brand gets recommended, because the assistant’s answers are increasingly shaped by what it already knows about the person asking. And it alters what a marketer’s own account is quietly accumulating — client names, draft strategy, account detail — every time the tool is used for real work.

What OpenAI changed

The new architecture, which OpenAI’s own evaluations label Dreaming V3, was announced in a post titled “Dreaming: Better memory for a more helpful ChatGPT”. Rather than a manually curated list of saved memories, a background process now reads across past conversations and decides what to retain, updating it over time. OpenAI’s example is telling: a note that a user “is going to Singapore in July” rewrites itself to “went to Singapore in July 2026” once the trip has passed, with no prompt from the user.

Three controls ship alongside it: a readable summary page showing what the assistant has synthesised, the ability to add or correct details, and settings for which topics it should raise and when. The company says the summary may not show every detail used to personalise a response. The rollout began with Plus and Pro users in the United States, who also get roughly double the memory capacity, before extending on 9 June to Go and Free users, whose responses draw on a smaller window of history. OpenAI credits a large cut in the compute needed to run the feature for making the free-tier expansion practical.

OpenAI reports steep gains on its own benchmarks — factual recall rising from 41.5% in its 2024 system to 82.8% with Dreaming V3, and a test of whether the assistant updates stale context climbing from 9.4% to 75.1% over the same period. These are the company’s figures, not independently verified, and saved memories themselves only launched in April 2024, so the trend matters more than the decimal points.

What it means for getting found

The visibility story is the personalisation. When a buyer asks ChatGPT to recommend a tool, the answer is no longer a single ranked list served to everyone; it is shaped by what the assistant has learned about that person’s role, preferences and past projects. The same question can produce different recommendations for different users, and the marketer cannot see, let alone tune for, any individual’s stored context.

That makes the discipline less about chasing a universal answer and more about durable authority that holds up across many personalised versions of the same query. A brand that is consistently and credibly present in the sources these systems trust is the one likely to surface whatever the user’s history nudges the assistant toward. Personalisation, in other words, raises the value of genuine reputation and lowers the payoff of anything that looks like gaming a single result. It is the same lesson the generative-search shift has been teaching, sharpened: you are optimising to be the brand an assistant is comfortable recommending to someone it thinks it knows.

What it means for marketers’ own data

The flip side is closer to home. Marketing teams now run a great deal through ChatGPT — client names, positioning, draft campaigns, sometimes a connected inbox — and a system that persistently synthesises and retains that across sessions changes the risk profile, especially on free, unmanaged accounts that until recently kept nothing. Early users have reported “context bleed”, where a detail from one conversation surfaces in an unrelated one, and the audit trail is incomplete by OpenAI’s own account. The wider backdrop is not reassuring: a US class action filed in May 2026 alleges ChatGPT embedded third-party tracking code, and an Italian regulator fined OpenAI €15m in late 2024 over data processing.

None of this argues against using the tool, which is plainly useful. It argues for treating it like any system that now keeps a record. The practical moves are unglamorous: know which tier and account each task runs on, use Temporary Chats or switch memory off for anything client-confidential, review the memory summary periodically to see what has been retained, and set workspace controls before a team pours sensitive work into personal logins. The discipline is the same one good marketers already apply to a CRM or an analytics tag — assume the data persists, and decide on purpose what goes in.

Memory is becoming a shared foundation across ChatGPT’s plans, and Claude and Gemini run their own versions of the idea. The assistant that remembers is both more useful and more consequential, and for marketers the single upgrade lands as two jobs at once: earn a place in a discovery surface that now personalises its answers, and govern a data surface that no longer forgets.