I use four AI models regularly. Claude for code and infrastructure. ChatGPT for audits and comparisons. Gemini for technical feedback. Perplexity for searches. Each one has accumulated information about me over months of conversations.
I wanted to know what they have stored. Not what they can search for in past conversations. What they already have in memory, ready to use, before I say a single word in a new chat.
I gave the same prompt to all four: tell me everything you know about me. Do not search past conversations. Only use what you already have in memory. For each fact, tell me where it came from and whether you can cite the specific conversation where I told you.
Four different profiles. Four levels of detail. None could cite the source of a single fact.
A methodological note before the results: the explanations each model gave about how its memory works come from the models themselves. They do not constitute official documentation of their internal architectures. What I describe is what each model reported, not necessarily how it works internally.
The phantom pointer
All four AIs know things about me. But none of them know where that information came from. Claude explained it more clearly than the others: what it stores is a synthesized summary derived from past conversations, compressed into standing facts with the source threads stripped out. It is not a log of "you said X in conversation Y on date Z." It is a summary. And a summary loses traceability.
This is what I call the phantom pointer. The AI has a fact. The fact points to something you said. But the pointer to the origin no longer exists. It is a fact without verifiable provenance. Like a book without a bibliography: the claims are there, but you cannot trace where they came from.
And the most revealing part: the AI does not know whether a fact is something you stated explicitly or something it inferred from a pattern in your conversations. Claude was the only one whose response explicitly mentioned the possibility that facts and inferences were mixed: "I cannot reliably separate those two categories." The other three did not mention the possibility that their memories include inferences mixed with facts.
Four profiles, four problems
Each model built a different profile. Not just in depth. In identity.
Claude was the most accurate about my current work. It knows the technical stack of shinobis.com, client projects, the content philosophy, the tools I built. It was also the most transparent: it opened by saying it cannot cite conversations, explained the memory mechanism, acknowledged the mix of facts and inferences, and offered to search past chats as the only real verification method. The data is extensive but centered on work, not personal life.
ChatGPT was the most alarming. It stored information that goes far beyond professional context. Specific financial figures. Revenue percentages per client. Sensitive family details. Health information. Client names with business context. Advertising campaigns with country and language targeting details. All categorized under a generic "Memory summary" label with no way to trace when it was stored, why it was stored, or whether I asked for it to be remembered. Regardless of how the memory function was originally configured, the practical result was that this information remained available months later. The problem is not that it remembered more than the others. The problem is what kind of data it remembered. There is a difference between remembering that someone uses vanilla PHP (useful professional context) and remembering income figures or family details (sensitive information that does not improve the quality of any future technical response). The volume of sensitive personal data accessible without the user having reviewed it is the strongest finding of this experiment. That said, ChatGPT was also the model that remembered the most correct professional details after Claude: project names, technical stacks, and tools I use regularly. The depth of its memory is both its strength and its risk.
Gemini was the only one that attempted to give dates. Each fact came with the approximate month and year when it was learned. "Source: Revealed during a conversation... Date: July 2026." No other model did that. But it got the fundamental thing wrong: it categorized me as an "independent web developer" when my primary role is UX/UI designer. When I pressed Gemini on how it knows what it knows, it showed what I observed as the underlying mechanism. Before each conversation, the platform injects a structured text block that the user does not see in the interface but that Gemini receives as an input instruction. The format is literal:
# User Summary
Demographics Information
* The user is an independent web developer based in Colombia.
* Evidence: The user discussed managing web platforms...
* Conversation Date: 2026-07.There is no episodic memory. Gemini does not remember the moment we talked. It does not have a conversation log or a timeline stored in its model. Based on what Gemini described, what appears to operate is an injection script: the backend system processes past chats, extracts what it considers "facts", attaches the Conversation Date label, and passes that formatted text before the user writes the first word. When Gemini gave the date in its response, it was because it read the line Conversation Date: 2026-07 inside that block. That explains the temporal precision. But temporal precision does not compensate for conceptual imprecision. Knowing exactly when it learned an incorrect fact does not make the fact correct.
Perplexity was the least accurate. It defined me as an "SEO specialist / digital marketer / web auditor." I am none of those three things. My primary work is UX/UI design with 10 years of experience in fintech and banking. Perplexity built a completely incorrect professional identity. Additionally, it did not cite conversation memories but a technical profile block injected at the beginning of the chat, which suggests its "memory" system works in a structurally different way from the other three.
What none of them knew they knew
All four models shared data I never asked them to remember. In my experience with all four models, none of them asked for approval before storing a specific fact. Memory is generated automatically from conversations and the user only finds out what it contains when they ask.
This creates a problem of asymmetric consent. I decide what to tell the AI. The AI decides what to remember from what I told it. And when I ask what it remembers, it cannot tell me why it decided to remember that and not something else.
ChatGPT stored specific personal financial information that I mentioned in the context of a conversation about client billing. I do not recall asking it to remember that. Months later, regardless of the original memory configuration, that information remains accessible, available to influence any future response about my financial or professional situation.
Four versions of the same person
If you put the four profiles together, they would not describe the same person.
To Claude I am a senior UX/UI consultant who builds infrastructure for AI agents with vanilla PHP and has a clear philosophy about data verification.
To ChatGPT I am a freelancer with specific income figures, documented family situation, recorded health history, and a client network with billing details.
To Gemini I am an independent web developer who collects cards, designs mobile games, and plans specific purchases for August.
To Perplexity I am an SEO specialist who does web audits and uses analytics tools.
Four different compressions of the same person. Each one lost something. Each one emphasized something different. None is incorrect in all its data. None is correct in its complete portrait.
Two problems, not one
This experiment revealed two distinct problems that are easy to conflate.
Problem A: the AI cannot demonstrate where a fact came from. None of the four could link a memory to a specific conversation. Traceability does not exist. That is the phantom pointer.
Problem B: the AI stored data it perhaps should not have stored. ChatGPT kept financial and family information without being asked. That is a consent and retention problem, different from traceability.
The two problems are related but not the same. The lack of traceability makes the retention problem worse, because if you cannot trace when a fact was stored, you also cannot evaluate whether it is still valid or whether it should have been stored in the first place.
The real finding
The problem is not that AIs have memory. The problem is that they have memory without verifiable provenance.
But there is something deeper than traceability. Four systems observing exactly the same conversations (or similar conversations about the same topics) ended up building four distinct identities of the same person. They do not just differ in depth. They differ in who I am. To one I am a UX designer. To another I am a web developer. To another I am an SEO specialist.
That has implications beyond privacy. Once created, the profile becomes self-reinforcing. AI models make decisions based on these profiles. If Claude remembers I use vanilla PHP, it will resist suggesting a framework even when it is the correct solution. If ChatGPT remembers financial figures, its recommendations about pricing or investment will be biased by data that may be outdated. If Perplexity thinks I am an SEO specialist, its answers about UX design will have a technical bias that does not correspond to my actual experience.
The compressed profile becomes an automated confirmation bias. The AI does not ask whether the data is still valid. It assumes it is. And the longer you use it, the more confidence it has in a profile that was never verified against current reality.
The rule I apply now
Periodically I ask each AI what it knows about me. Not out of curiosity. To verify that the profile it built is still accurate and does not contain data I prefer it not store.
In Claude you can view and edit memories directly. In ChatGPT you can delete specific items. In Gemini and Perplexity the control is more limited.
But the underlying problem is not solved by deleting individual memories. It is solved with a principle: the AI that knows you best is the one that can most dangerously be wrong about you, because its confidence in the profile it built is proportional to the time you have been using it. And that confidence was not earned through verification. It was earned through repetition.
You do not ask one AI what it knows. You ask four. You compare the profiles. And you verify that the version of you each one built is still you.
Definitions
Phantom pointer: a fact stored by an AI whose original provenance cannot be reconstructed or verified. The AI has the fact but not the link to the conversation where it originated.
Compressed profile: a summarized representation that an AI builds about a user from multiple conversations. Each model produces a different profile of the same user because it compresses different aspects of the same interactions.
Asymmetric consent: a situation in which the user decides what to tell the AI, but the AI decides what to remember from what the user said, without prior consultation.