Open the Memory settings inside ChatGPT and a list appears — preferences, past corrections, facts the model has inferred about you across sessions — each one editable or deletable individually. It is a small interface for an enormous decision: what an AI system is allowed to consider permanently true about a person, made by a product team rather than a court or a regulator.
Memory systems have to make constant, mostly invisible decisions about salience — what counts as worth remembering, what decays, what gets summarized versus stored verbatim. Those decisions determine what the system later treats as established fact about a person, which is functionally an editorial judgment even when no human reviews it.
More persistent memory makes an assistant more useful by removing the need to repeat context. It also makes an early misunderstanding sticky — a wrong inference about someone's intent or identity, once stored, can quietly shape every subsequent interaction until something forces a correction.
The stakes compound at population scale in ways individual memory failures never did. OpenAI's, Google's, and Anthropic's retention choices, replicated across hundreds of millions of users, function like an editorial policy for what an entire society's AI layer considers relevant about people — not unlike how search-indexing choices shaped two decades of what counted as findable.
The three approaches already diverge in instructive ways. ChatGPT's memory is explicit and user-editable by default. Google's Gemini personalization draws more heavily on account-wide signals like search and Workspace history. Anthropic has moved more cautiously, scoping Claude's persistent memory to individual Projects rather than a global profile. None of the three has been forced to justify that design choice to an outside body.
GDPR's right to erasure was built around discrete data requests and deletions, not around a system that continuously reweights what it has learned about someone in ways that are hard to audit or fully reverse. That makes 'delete my data' a much murkier promise for a memory-based assistant than for a conventional database, and regulators have not yet caught up to the distinction.
The products that earn durable trust will likely be the ones that make memory legible — letting users see, edit, and correct what has been retained the way ChatGPT's Memory panel already attempts — rather than the ones that simply retain more. Legibility, not volume, is what turns memory from a liability into an asset.
Memory used to be the thing AI systems lacked. Now that OpenAI, Google, and Anthropic each have some version of it, each has also quietly decided, on its own, how much of it a user is allowed to see — and none of the three has had to defend that choice to anyone outside the company yet.