deep-dive

Cross-Platform AI Memory: Who Controls What You Remember

AI assistants now build persistent memories about you. As those memories start moving between platforms, privacy gets harder to track and far easier to lose.

For most of computing history, software had no memory between sessions. You opened a program, used it, closed it, and next time you opened it, it had no recollection of anything that had happened before. Each session started fresh.

AI assistants have changed that. ChatGPT, Claude, Gemini, and dozens of smaller tools now maintain memory across sessions — remembering your preferences, the projects you are working on, your work habits, your health concerns, the names of your family members, and the ongoing texture of your life. When you tell an AI something, that something is now potentially stored and recalled indefinitely.

This is useful. It is also the beginning of a privacy challenge that is only beginning to come into focus: what happens as these memories start moving between platforms?


What AI Memory Actually Stores

The specifics vary by platform, but modern AI memory systems generally store some combination of:

Explicit facts you share. “I’m a software engineer.” “I have two kids.” “I’m vegetarian.” When you share these directly, they are stored and recalled in future conversations.

Inferences from your behaviour. An AI that observes many conversations may infer things you have never explicitly stated: your approximate level of technical knowledge, your general political leanings, your recurring anxieties, your communication style. Some systems store these inferences as memories.

Conversation context. Beyond facts, some systems store summaries of past conversations — what you were working on, what problems you described, what advice you received. This builds something closer to a relationship history.

Habits and patterns. Systems that integrate with your calendar, files, email, or other services can build a richer picture: when you typically work, what projects recur, which contacts you interact with frequently.

The cumulative effect across months or years of regular use is a detailed model of who you are — constructed not from what you have deliberately disclosed, but from the aggregate of everything you have shared, inferred, and implied.


The Single-Platform Problem

Each major AI platform currently operates its own memory system. ChatGPT memories stay on OpenAI’s servers. Gemini memories stay on Google’s. Claude’s memory context stays within Anthropic’s infrastructure.

This is both a privacy feature and a limitation.

As a privacy feature: your memories on one platform are not automatically accessible to another. If you switch from one assistant to another, the new service does not inherit the accumulated context from the old one.

As a limitation: you cannot easily take your memories with you. The model you have trained on your preferences, the context accumulated across hundreds of conversations, the relationship history you have built — it stays with the platform. Switching assistants means starting over, which creates real switching costs and reinforces lock-in.

A researcher at MacPaw published a preprint in mid-2026 arguing that AI memory needs an open portable format, citing this lock-in as a fundamental problem for user autonomy. A workshop at Stanford’s HAI group in the same period described the risk of personal context concentrating within proprietary “cognitive layers” controlled by large technology companies.


The Push for Portable Memory

The emerging solution — and the emerging privacy risk — is cross-platform AI memory.

Several vendors are building memory systems explicitly designed to sit outside any single AI platform, storing a user’s accumulated context in a centralised layer that multiple AI tools can access. The selling point: your memories travel with you. Switch from one assistant to another, and your preferences, history, and context are immediately available to the new tool.

The privacy implication: instead of your memories being siloed across separate platforms with separate privacy terms, they are now held by a third party whose terms govern all of them. When you export memories from Platform A and import them to Platform B, the terms governing those memories change with the destination.

The specific risks this creates:

Privacy policy arbitrage. A memory that was created under a platform with strong privacy terms may end up in a platform with weaker ones if the user moves their memory export.

Aggregation across contexts. Memories built from different types of interactions — work tasks, health questions, personal conversations, financial planning — may have been generated in contexts where you expected them to be isolated. A unified cross-platform memory removes that isolation.

Breach amplification. A single memory provider that holds context from all your AI interactions becomes a highly attractive target. A breach of that provider exposes everything, rather than isolating the damage to one platform’s memory.

The question of what is portable. If AI memory is portable, it may also be shareable, compelled, or sold in ways that individual conversation histories cannot be. The legal frameworks for personal data portability do not necessarily create the privacy protections that individual data would receive.


Google’s Telling Hesitation

One of the clearest signals that cross-service AI memory raises genuine regulatory concerns came from Google’s own decisions about deployment.

Google’s Personal Intelligence feature — which connects Gemini AI to Gmail, Google Drive, Google Photos, Google Calendar, and Google Maps simultaneously — was held back from EU, UK, and Swiss users when it launched. The service, which allows Gemini to synthesise personal context across all of Google’s major products into unified, responsive AI assistance, faced scrutiny from EU regulators over whether connecting this breadth of personal data sources for AI processing was compatible with GDPR.

The decision to restrict the feature geographically is a data point about what regulators in those jurisdictions think about unified cross-service AI memory.

Google’s unified approach represents the logical endpoint of cross-platform memory within a single ecosystem: one AI that can see everything you have ever stored, done, or communicated through Google’s products. The privacy implications of that model are part of why the EU and UK have treated it differently from less integrated AI features.


What Controls Actually Exist

For the platforms that offer memory today, controls are inconsistent.

ChatGPT allows users to view, edit, and delete stored memories through the memory management interface. Disabling memory stops new memories from being stored and clears existing ones. The control is real but not prominently surfaced.

Gemini has memory features tied to Google’s broader account activity — search history, Gmail summaries, and connected service data. Controls are distributed across different settings pages rather than unified in one place.

Claude as of this writing uses conversation context rather than a persistent memory store that persists across separate conversations by default, though this may change with model updates and product evolution.

Third-party memory layers vary significantly. Some offer transparent audit logs and deletion controls. Others treat accumulated context as proprietary data with limited user access.

The consistent challenge is that memory accumulates passively. You do not opt into each individual memory. You make one setting choice early on, the system observes everything, and memories are created from that observation. Reviewing what has been stored requires active effort — opening a settings page most users will never visit.


The Informed User’s Position

There is no single right answer to how much memory you should give an AI assistant. A heavily personalized AI is genuinely more useful for many tasks. The accumulated context from months of interaction means less repetition, more accurate assistance, and faster task completion.

The trade-offs worth understanding:

Stored memories are stored data. They are subject to your platform’s privacy policy, their retention practices, and their legal obligations to disclose data under court orders or regulatory requests. A memory that tells an AI you have been dealing with a health condition, or that you are stressed about a legal matter, is data that now exists on a third-party server.

Deleting a conversation does not delete memories created from it. On most platforms, memories persist until explicitly deleted. Clearing your conversation history leaves the memories intact.

Memory exports are potentially sensitive. If you use a memory portability tool and export your AI memories, that export file may contain some of the most sensitive personal information you have shared with an AI. It deserves the same protection as any sensitive personal file.

On-device memory is different from cloud-stored memory. Some local AI tools — those running on your computer or phone without a cloud component — keep their memory local. This is a genuinely different privacy model.


Storing What Matters Separately

The practical implication for people who want to maintain their own record of personal context — their own second brain, their own memory archive — is that this record should live somewhere they control, not on platforms whose terms can change.

The difference between AI memory and personal file storage is significant: when you store a document, photo, or note in private storage, you decide what is there, you can review it, and you can delete it. When an AI builds a memory model from your interactions, what gets stored is determined by the platform’s inference engine, not by your choices.

Keeping a personal archive of notes, documents, photos, and journal entries in storage you control, encrypted without analysis, means that your personal context is not dependent on any AI platform’s decisions about what to remember and what to forget.

The emerging cross-platform AI memory infrastructure is an answer to real limitations. It is worth watching carefully — because what it promises in convenience, it costs in the kind of control that most people have not yet had reason to think carefully about.

The question is not whether persistent AI memory is useful. It clearly is. The question is whether the version that gets built will be one where you understand and control what is remembered about you, or one where the memory belongs to the platform.

Your memories deserve better than an ad platform.

Try daftei free →
← All posts