Five Companies, Five Memory Systems, Zero Interoperability: What ChatGPT, Claude, and Gemini Actually Store in Memory About You

Every major AI assistant now builds a persistent memory of who you are. Within a single quarter of 2026, three distinct architectures shipped or expanded: one that synthesizes your profile autonomously while you sleep, one that makes its memory legible and editable, and one that reads your email, calendar, and photos to assemble a personal intelligence layer. They all call it "memory." They mean completely different things by it. And none of them talk to each other.

Key Takeaways

• OpenAI's Dreaming V3, Claude's free-tier memory, and Gemini's Personal Intelligence all shipped or expanded between March and June 2026, each with a fundamentally different architecture for what gets stored, how it updates, and whether you can inspect it.
• No structured portability standard exists. Every "import" tool currently available is a copy-paste prompt hack, not a real data export. Switching providers means losing your memory or reconstructing it manually.
• The EU AI Act's Article 50 transparency rules take effect August 2, 2026, but they mandate disclosure ("you're talking to a machine"), not memory portability or auditability. The law solved the wrong problem.
• Gemini's Personal Intelligence is deliberately withheld from the EEA, UK, and Switzerland, which suggests the most aggressive memory architectures can't comfortably ship where privacy regulation is strictest.
• We built Selina's memory to be user-owned, auditable, and encrypted at rest, but we are honest about its limits: memory is not end-to-end encrypted, because a slice of each request reaches a frontier provider at inference.

Why Did Three Memory Systems Ship in the Same Quarter?

Because the product ceiling for stateless chatbots was reached sometime in 2025. A model that forgets you between sessions is a demo. A model that remembers you is a product. Every major provider arrived at this conclusion independently, and the result is a three-month pile-up of incompatible memory architectures landing right before the EU's August 2 transparency deadline.

Anthropic dropped the memory paywall for free Claude users on March 2, 2026, after first introducing memory as a paid feature in August 2025 and expanding it to all paid subscribers by October. OpenAI began rolling out Dreaming V3 on June 4, 2026, replacing the old manually curated saved-memories list with an autonomous background synthesis process. Google's Personal Intelligence layer for Gemini expanded through the same window, connecting chat memory to Gmail, Calendar, Photos, and Drive.

Three architectures, three threat models, three definitions of what "your memory" even means. Zero interoperability between any of them.

What Does OpenAI's Dreaming V3 Actually Store?

Dreaming V3 replaced the explicit, user-editable memory list with an autonomous background process that builds a model of who you are from your entire conversation history. The system runs asynchronously. You do not trigger it. OpenAI's own example illustrates the temporal-rewrite behavior: a memory like "the user is going to Singapore in July" rewrites itself to "the user went to Singapore in July 2026" after the trip ends, with no user action required.

The capability gain is real. OpenAI reports factual recall rising from 41.5% in 2024 to 82.8% in 2026 on its own internal eval, with preference and time-sensitive scores landing in the low-to-mid 70s. The company describes these numbers as "directional rather than verified," which is an unusually honest qualifier for a benchmark you designed yourself.

The tradeoff is auditability. The explicit memory list had a useful property for enterprise IT teams: it was auditable. You could see exactly what the model had saved, edit individual entries, delete them. The background synthesis approach is more opaque. The model decides what matters. It rewrites entries over time. You can still view and delete memories, but the process that generates them is no longer fully transparent.

For individual users, this might feel like a better product. For compliance teams, it is a step backward.

How Does Claude's Memory Differ Architecturally?

Claude's memory was explicitly built around what Anthropic calls "legibility": the idea that users should be able to see and audit what the model knows about them. When Claude gained its first memory capability on August 11, 2025, it was an explicit, opt-in recall feature rather than a background profile. That philosophy carried through to the free-tier rollout.

The practical difference: Claude's memories are discrete entries, each created from a specific conversation, each individually viewable and deletable. The model does not autonomously synthesize a profile from your history while you are offline. It remembers what you told it to remember, or what it asked permission to remember.

Anthropic reports free-plan users are up 60% since the start of 2026, with paid Pro and Max subscribers doubling. Some of that growth coincides with a Pentagon contract dispute that put Anthropic in headlines for other reasons, but the memory rollout was the product story.

The most interesting artifact of the rollout is the migration tool. Anthropic built a competitive import tool that works by having users copy a specific prompt into a competing AI chatbot. The prompt instructs the rival system to export all stored memories and context, formatted as structured entries. It is a clever hack. It is also, architecturally, a confession: there is no standard export format, no API, no protocol. You are literally asking one chatbot to dictate its memories to another via clipboard.

What Makes Gemini's Personal Intelligence Different From the Others?

Gemini's Personal Intelligence is not just chat memory. It is a cross-application data layer. The feature lets the assistant read across connected Google apps, including Gmail, Calendar, Photos, and Drive, together with its memory of past chats. It is opt-in and off by default.

The memory storage itself takes a structured, schema-based form. Gemini stores what it knows about you in a single document called user_context, a typed outline made up of short, factual bullets in a fixed schema with defined sections. This is more structured than the free-text approach other providers use, and arguably more auditable at the individual-entry level, though the cross-app data ingestion makes the total surface area much larger.

Gemini's internal rules restrict inference on protected categories, including mental or physical health conditions, national origin, racial or ethnic origin, religious beliefs, sexual orientation, and political opinions. This is a policy guardrail, not a technical one, and competitors have not all published equivalent restrictions.

Google separately renamed its older "past chats" feature to "memories," folding chat history and connected-app access into a single personalization narrative. The rebrand is worth noting because it blurs the line between "things you said to the chatbot" and "things the chatbot read from your inbox." Those are very different categories of data with very different user expectations.

Why Is Gemini's Most Aggressive Memory Feature Missing From Europe?

Because Google appears to be gating personal data features precisely where privacy regulation is strictest. Personal Intelligence is currently unavailable in the European Economic Area, Switzerland, the United Kingdom, and Nigeria.

The timing is hard to ignore. The EU AI Act's Article 50 transparency obligations take effect August 2, 2026, requiring every chatbot and AI agent to disclose to users that they are interacting with a machine. The EU's July 2026 "Digital Omnibus" delayed other AI Act provisions but left Article 50 intact.

If your memory architecture cannot comfortably ship in the jurisdiction with the strongest privacy framework, that tells you something about its actual risk profile. It is not proof of wrongdoing. It is a revealed preference about where the compliance friction is too high to bother.

What Does the EU AI Act Actually Require for Memory?

Almost nothing, specifically. Article 50 requires chatbots to tell people they are talking to a machine, with fines reaching €15 million or 3% of global revenue for companies that ignore it. That is the transparency obligation that survived the omnibus reshuffle.

What Article 50 does not require: that a chatbot tell you what it remembers about you, that you can export your memory in a portable format, that inferences drawn from your conversations are auditable, or that you can move your memory to a competing service.

GDPR's existing right to data portability theoretically covers some of this ground, but in practice, no chatbot provider offers a structured, machine-readable memory export. The "import" tools from Anthropic and Google are prompt-based workarounds that rely on the source chatbot cooperating with a natural-language request to list its memories. That is not portability. That is social engineering your own AI.

The regulatory gap is precise: disclosure was mandated, but portability was not. You will know you are talking to a machine. You will not necessarily know what it thinks it knows about you, or be able to take that knowledge elsewhere.

How Do These Memory Architectures Compare on Auditability?

The spectrum runs from fully opaque to mostly legible, with no provider achieving full transparency.

OpenAI's Dreaming V3 is the most opaque. The synthesis runs autonomously. Entries rewrite themselves over time. You can view the result but not the reasoning that produced it. Enterprise IT teams have flagged the auditability regression as a concern for compliance workflows that previously relied on the explicit memory list.

Claude's approach is the most legible among the major providers. Discrete entries, traceable to conversations, individually deletable. The design reflects Anthropic's stated emphasis on bounded context and user inspection. The limitation is capacity: legible memory does not scale as aggressively as autonomous synthesis, and the import tool's reliance on prompt-based extraction means the migration path is fragile.

Gemini's structured schema (user_context) is auditable at the storage layer, but the cross-app data ingestion means the inputs to that schema are vast. The sensitive-attribute guardrails are a policy layer on top, not a structural guarantee. And the feature's absence from Europe suggests the auditability story is not yet strong enough for the strictest regulators.

What Does "Zero Interoperability" Actually Mean for Users?

It means your relationship with an AI assistant is a lock-in mechanism. Every conversation that builds memory is an investment in a proprietary profile you cannot take with you. Switch from one provider to another and you start from zero, unless you use a clipboard-based import hack that may or may not capture everything.

There is no Memory Transfer Protocol. No standard schema. No API endpoint where you can request a structured export of what an AI thinks it knows about you. Anthropic's import tool instructs users to paste a prompt into a competing chatbot requesting an export of all stored memories, formatted as dated entries in a code block. This works only if the competing chatbot cooperates, only if the user copies the output correctly, and only for memories that the source system surfaces in response to a natural-language request.

The parallel to early email or social media is obvious. Before IMAP, your email was trapped in whichever webmail client you chose. Before data portability mandates, your social graph was a proprietary asset. AI memory is at that same pre-standard stage, except the data is more intimate than your inbox or your friend list. It is a model of how you think.

What Should a Memory Architecture Actually Look Like?

We have opinions on this, because we built one. Our position is that memory should be user-owned, auditable entry by entry, and encrypted at rest. We also think you should be able to delete it and have "delete" mean actually gone.

We are honest about the limits. Selina's memory is encrypted at rest, but it is not end-to-end encrypted. A slice of each request reaches a frontier provider at inference. That is the architectural reality of any system that routes to external models for generation. We use a stack of frontier models, routed per task, and the memory context that accompanies a request is exposed to the provider handling that request. Files and transfers through SelinaSEND are zero-knowledge encrypted. Memory is not. We state that plainly because the distinction matters.

Non-content operational metadata is kept for a short retention window, not zero retention. The account itself is protected; the content within it is encrypted. These are different claims and we do not blur them.

What we do claim: memory is adaptive, not a transcript. It is not a verbatim log of everything you have ever said. It is a structured representation that updates as your context changes. You can see every entry. You can delete any entry. Delete means gone. Actually gone.

We also think a memory architecture should ship the same way everywhere. The same feature set in the EU as in the US. If your architecture requires you to withdraw from an entire jurisdiction because the privacy rules are too strict, that is not a regulatory inconvenience. That is a design flaw.

Why Does "Disclosure Without Portability" Miss the Point?

Because knowing you are talking to a machine is table stakes. The harder question is whether you can see what the machine has concluded about you, correct it, and take it with you if you leave.

Article 50 of the EU AI Act addresses the first problem. No current regulation meaningfully addresses the second. GDPR's right to data portability was designed for structured databases, not for probabilistic inferences drawn from conversational history and synthesized into a living profile.

The result is a regulatory framework that will, as of August 2, 2026, require your chatbot to identify itself as artificial, while placing no constraints on how it builds, rewrites, or silently expands its internal model of who you are. The chatbot must say "I am not human." It does not have to say "here is what I think I know about you, and here is how I arrived at it."

A right to memory portability, analogous to data portability but specific to AI-constructed user models, would require a standard export format, a standard schema, and a regulatory mandate. None of those exist. The copy-paste prompt hacks are the industry's current answer, and they are not an answer.

What Happens Next?

Memory becomes the primary competitive moat for AI assistants in 2026 and 2027. The models themselves are converging in capability. The differentiator shifts to context: who knows you best, who remembers most accurately, who can act on what they know across the most surfaces.

That makes the interoperability problem worse, not better. Every provider is incentivized to deepen lock-in through richer memory, not to standardize it. Anthropic's import tool is a customer-acquisition wedge, not a portability standard. Google's cross-app integration makes the switching cost astronomical for anyone deep in the Google ecosystem.

The regulatory window is narrow. Article 50's transparency rules are a floor, not a ceiling. If memory portability is going to happen, it will need to be mandated the way data portability was. The alternative is what we have now: five companies, five memory systems, five proprietary profiles of who you are, and a clipboard as the only bridge between them.

We built Selina to hold memory differently. Not perfectly. Not with guarantees we cannot make. But with a design that assumes you will want to see what is stored, delete what you do not want, and know exactly what is and is not encrypted. If that sounds like the kind of architecture you want to try, start a free 7-day trial, no card required.

Michael C.

Founder & Principal Engineer, Selina Labs

Michael builds Selina, a privacy-first AI that remembers you across conversations. He ships security-sensitive AI in production — real attacks, real fixes, measured in minutes and dollars — and writes about privacy, security, and LLMs from that seat. Top Rated Plus and expert-verified on Upwork.

• www.upwork.com
• x.com
• medium.com
• selinalabs.io