ChatGPT memory now profiles users by work and habits

OpenAI's "Dreaming" memory system now keeps user information current 75.1% of the time, up from 52.2% a year ago. The Decoder reports that ChatGPT has shifted from saving scattered bullet points to assembling coherent narrative profiles, sorted into categories such as work, hobbies, and travel. The model is no longer a stateless oracle answering each prompt fresh. It is a persistent reader of you.

That shift changes the unit of competition for brand visibility. Until now, the prompt was the battlefield: a user typed a query, the model retrieved and ranked sources, a handful of brands got cited. Memory introduces a second, slower battlefield: the dossier. What the model already believes about a user, including which vendors, advisors, frameworks and institutions belong in their professional world, now shapes which brands surface before retrieval even fires.

Consider a procurement director at a European utility who has spent six months asking ChatGPT about cement decarbonisation, Scope 3 reporting, and ISO 14068. The next time she asks "who should I shortlist for a low-carbon binder pilot?", the model is not starting cold. It has a work profile. It knows her remit. The brands embedded in that profile, named in prior answers, saved across sessions, get a structural advantage that no amount of fresh SEO can dislodge in the moment.

This is the mechanism CMOs should understand: memory compounds citations. A brand cited once in a user's earlier session is more likely to be cited again, because the model has filed it under that user's domain of interest. Visibility becomes path-dependent. The first mover in a user's dossier wins disproportionately, in the same way the first analyst a banker calls tends to become the analyst they always call.

The 75.1% figure matters because it crosses a threshold of reliability. At 52%, memory was a novelty feature; users could not trust it and neither could brands planning around it. At three-quarters accuracy, the dossier is durable enough to be treated as infrastructure. OpenAI is signalling that persistent user state is now a product, not an experiment.

Three implications follow for B2B brands whose buyers use ChatGPT at work.

First, the cost of being absent from early-stage conversations rises sharply. A bank that fails to be cited when a treasurer is exploring tokenised deposits in 2026 may find itself absent from that treasurer's dossier in 2027, when the actual RFP shapes up. Multilaterals face the same risk in reverse: an institution like the IMF or the World Bank that dominates a policymaker's early research sessions becomes the default reference for years of subsequent queries. Authority calcifies.

Second, the categories themselves are revealing. Work, hobbies, travel. OpenAI has chosen taxonomies that mirror how consumer products segment audiences, not how enterprise buyers think. "Work" is a single bucket. That means a CFO's queries about hedging, ERP migration, and audit risk land in one undifferentiated profile, and the brands the model associates with her professional identity will be drawn from across that whole surface. Narrow category dominance matters less than broad presence across a buyer's working life.

Third, memory raises the value of being the brand a user explicitly saves or names. ChatGPT lets users edit memories. Marketers should assume their most engaged prospects will, over time, curate the brands they want the model to remember. Earning that explicit mention, through product experience, useful content, or a memorable point of view, is now a measurable visibility asset. Generic thought leadership does not get saved to a dossier. Specific, named, useful frames do.

There is a regulatory shadow here that financial-services and UN-system communicators should track. Persistent user profiles built by a US vendor, sorted by behavioural category, sit awkwardly with GDPR's purpose-limitation principle and with the data-minimisation expectations of multilateral procurement. Expect compliance teams to push back on staff use of memory-enabled ChatGPT for sensitive work. That fragmentation, some buyers operating with full dossiers, some with memory disabled by policy, will produce uneven visibility patterns that marketers will need to measure separately.

The broader point: LLM visibility is no longer a single-shot ranking problem. It is becoming a relationship problem, mediated by a model that remembers. Brands that treat each ChatGPT answer as an isolated event will lose ground to those that understand they are being filed, by user, into a profile that quietly shapes every subsequent answer.