AI Memory Wars: How ChatGPT, Claude, and Gemini Remember You (And What It Means for Your Digital Self)

The Quiet Shift: AI Learned to Remember in 2024-2026

For most of ChatGPT's first year, every conversation started from zero. The model didn't know who you were, what you worked on, or what you'd asked five minutes earlier in another tab. That was the deal. It was also the reason privacy-conscious users trusted it more than, say, their search history.

That deal ended in February 2024. OpenAI announced Memory as a research preview. By September 2024 it was generally available. On April 10, 2025, OpenAI upgraded the feature so ChatGPT could reference all past conversations, not just items you explicitly asked it to remember. Free-tier rollout followed on June 3, 2025. A tool used by hundreds of millions of people had become persistent almost overnight, and most users didn't change a single setting.

Anthropic took the opposite pace. Claude Memory launched for Team and Enterprise plans in September 2025, reached Pro and Max users in October 2025, and finally arrived on the free tier on March 2, 2026. Anthropic's launch blog emphasized something competitors had sidestepped: memory that users could actually read, edit, and delete.

Google shipped "Saved Info" in Gemini in February 2025 as a manual feature. In August 2025 it flipped the switch: "Personal Context" became automatic and on by default for most accounts. Under-18 users, Workspace accounts, and anyone in the EEA, UK, or Switzerland were excluded, a pattern that tells you everything about which memory defaults survive European regulatory scrutiny.

The shift was quiet because each step was incremental. Taken together, it's one of the more significant privacy inflection points of the decade.

How Each System Actually Stores You

The marketing language is nearly identical. "Memory." "Personal context." "Learns about you over time." What varies is the underlying architecture, and that difference matters.

ChatGPT operates two memory layers. The first is "Saved Memories," which is explicit: you or the model flag a fact, and it's stored as a discrete item. The second is the chat history layer added in April 2025. This one is implicit. The model scans past conversations to extract patterns, preferences, and context, and feeds those signals into new chats. You can see saved memories in Settings, but the chat history layer is harder to audit directly. It's reasoning over text, not a list.

Claude took a different path. Its memory system is built around markdown files. On Claude Code and Claude.ai, long-term context lives in CLAUDE.md style files that you can open, read, edit, or delete line by line. The file approach means there's no black box. If Claude "remembers" that you prefer concise answers, it's because a literal sentence to that effect sits in a file you can view. Claude also offers an "Incognito Chat" mode for sessions that shouldn't touch memory at all.

Gemini's Personal Context runs closest to ChatGPT's implicit layer. It automatically builds a profile from conversations across the Google ecosystem. The feature interacts with Gemini Gems, Google's custom persona system, so memories can be scoped to specific workflows. Temporary chats auto-delete after 72 hours, which is both a privacy feature and a reminder that "temporary" in AI products now has a specific shelf life.

Here's the feature matrix side by side:

The table hides one thing the official docs don't emphasize: defaults are destiny, and most users never change them.

The Transparency Gap

Claude's file-based approach is worth sitting with for a moment, because it reframes what memory is supposed to be.

If you ask ChatGPT what it remembers, you get a summary. If you ask Gemini, you get a gentle description of "what I've picked up from our conversations." If you ask Claude, you can open the actual file. The difference between a summary and a file is the difference between a verbal account of your medical history and the chart itself.

This matters for three reasons. First, audit. You can't meaningfully consent to a system you can't inspect. Markdown files give you line-item control: keep this, delete that, rewrite the third line. Second, correction. Implicit memory systems infer things, and inferences are often wrong. If ChatGPT quietly concludes you're a teacher because you asked about classroom policies once, that assumption shapes every future response and you have no way to see it. Third, portability. A human-readable file travels. A proprietary embedding does not.

Simon Willison put it bluntly on his blog on May 21, 2025: "I really don't like ChatGPT's new memory dossier." His complaint wasn't about privacy in the abstract. It was about opacity. Willison, who has spent years documenting LLM behavior, objected to a system that builds a model of him without giving him the model to inspect. He wasn't being paranoid. He was describing the core issue: AI memory creates a representation of you that you don't own, can't fully see, and can't cleanly move.

For context on why this shift matters beyond privacy theater, our piece on how to manage personal context in an AI-first workflow covers the operational side.

What Could Go Wrong: 2025-2026 Privacy Incidents

The abstract concerns became concrete fast. A short timeline of what actually happened:

The NYT order is the most structurally important. Users who explicitly deleted chats, including Temporary Chats that were never supposed to be saved, are now part of a legal hold they didn't consent to. Your right to delete AI memory is conditional on whether anyone is currently suing your AI provider.

The Chat & Ask AI leak is a different failure mode. A third-party wrapper using the underlying APIs misconfigured Firebase and exposed hundreds of millions of messages. The attack surface isn't just OpenAI, Anthropic, and Google. It's every startup layered on top of them.

Stanford's October 2025 analysis added the final piece. Even when providers intend to delete data, backups, safety review queues, legal holds, and training pipelines create retention windows that often exceed user expectations by orders of magnitude.

The Digital Self Question

Privacy framing only gets you so far. The deeper issue is what it means to be modeled.

When ChatGPT accumulates thousands of interactions with you, it isn't just storing facts. It's constructing a statistical approximation of how you think, what you care about, how you phrase things. That approximation then shapes every future response you get. The AI you talk to on day 300 is, in a meaningful sense, talking to a simulacrum of you while also talking to you. The loop is recursive.

Willison's "dossier" framing catches one side of this. The other side is psychological. Heavy AI users start to perform for how the system sees them. People adjust their prompts, their tone, even their self-descriptions to shape the model's impression of them.

This is different from how we perform for search engines or social media. Google doesn't talk back in your voice. AI does. And when the memory layer is opaque, you're performing for a mirror that shows a reflection you can't fully see.

Our earlier piece on the AI thinking trap covers the cognitive side of this. The memory layer adds an identity dimension. You aren't just at risk of outsourcing your thinking. You're at risk of outsourcing the authoritative record of who you are.

Memory Portability: March 2026 Changed Everything

For most of 2025, AI memory was a lock-in mechanism. The longer you used ChatGPT, the more it knew, and the more expensive it was to switch to Claude or Gemini. Memory was network effect compounded per user.

March 2026 cracked that. Anthropic launched an Import Memory tool that pulls context from ChatGPT and Gemini exports into Claude's markdown files. Google shipped its own cross-platform import around the same time. The question shifted from "how do I keep my context if I switch providers?" to "which provider gives me the best memory primitives, and how do I move between them?"

Two effects followed. Switching costs dropped for power users. And the surface area for abuse expanded: any tool that can ingest a memory export can reconstruct a detailed model of the user from it. Import Memory is a portability feature and a concentration-of-risk feature at the same time.

The interesting part is how this interacts with Claude's file-based approach. Because Claude memory is already human-readable markdown, you can bring files in, edit them, merge them, or write your own from scratch. That turns memory from something the AI does to you into something you do with the AI.

Practical: How to Audit and Control Your AI Memory Today

The defaults aren't going to protect you. Here's what to actually do, platform by platform.

ChatGPT

1. Open Settings, then Personalization, then Memory.
2. Review "Saved memories." Delete anything sensitive or inaccurate. Most users find at least one surprise here.
3. Toggle off "Reference chat history" if you want to limit the implicit layer. Note that this reduces personalization significantly.
4. Use "Temporary Chat" for anything you don't want referenced later. Be aware of the NYT retention order: temporary isn't the same as deleted.
5. Periodically export your data via Settings to audit what OpenAI holds.

Claude

1. Open your CLAUDE.md or equivalent memory file directly.
2. Edit it like any text file. Delete lines that are wrong or outdated.
3. Use Incognito Chat for sensitive sessions.
4. On Claude Code, your memory file is in your project repo. Keep it out of public commits unless you want your context indexed by every search engine.
5. Consider writing your own memory file from scratch: your preferences, your projects, your working style. It beats letting the model guess.

Gemini

1. Go to gemini.google.com, then Activity, then review saved info and activity history.
2. Toggle off Personal Context if you don't want automatic profiling. Expect reduced quality for personalized tasks.
3. Use temporary chats (72-hour auto-delete) for one-off sensitive queries.
4. If you're in the EEA, UK, or Switzerland, the feature is already disabled for you. This is worth knowing.
5. Check integration permissions: Personal Context can overlap with Gmail, Calendar, and Drive access. Audit those separately.

Across all three

• Don't paste anything into AI that you wouldn't put in a work email. Legal holds, leaks, and retention windows mean the delete button is aspirational.
• Treat AI memory as an extension of your public record, not a private notebook.
• Keep your real knowledge base somewhere you control. More on that below.

Toward User-Owned Memory

Here's the uncomfortable conclusion of the past two years. If memory is the layer that makes AI useful, then whoever owns the memory owns the relationship. Right now, for most users, that owner is OpenAI, Anthropic, or Google.

The alternative isn't to abandon AI. It's to own your own memory layer and let AI read from it, rather than the other way around.

This is where user-controlled tools fit in. Glasp's web highlighter lets you save passages, annotations, and notes from anything you read on the web, in a format you control, exportable at any time. Those highlights become a durable record of what you've engaged with and what you thought about it. They aren't inferred from your behavior by a vendor. They're authored by you.

Once you have that layer, Glasp's AI chat can read from it in a way you've explicitly opted into. You choose what goes in. You decide what to keep. You can delete a highlight and know that the deletion sticks, because the storage is yours. And because the highlights are text rather than opaque embeddings, they're portable to Claude, ChatGPT, Obsidian, or any future tool.

This isn't hypothetical. The Glasp community has been building personal highlight archives for years, often without realizing they were building the exact primitive the AI industry would later commodify and centralize. A well-maintained highlight library is the most practical form of AI memory available today, because you own it.

The broader frame is about legacy. We've argued elsewhere that your highlights and notes are arguably the greatest intellectual inheritance you can leave because they capture attention, taste, and judgment in a form that outlives the platforms that host them. The same piece of curation that makes a great AI context window also makes a great record of a mind. And in an era where AI will increasingly be the interface through which future generations read, the human curator role becomes more valuable, not less.

Frequently Asked Questions

Can I fully delete my ChatGPT memory?

You can delete saved memories and clear chat history through Settings. However, due to the 2025 NYT v. OpenAI court order, OpenAI is currently required to retain conversations that would otherwise be deleted, including some temporary chats. Treat AI memory as persistent until providers confirm otherwise.

Is Claude's markdown memory really safer than ChatGPT's?

It's more transparent, which is a different property from safer. You can read, audit, and edit Claude's memory in a way you cannot with ChatGPT or Gemini. Transparency makes it easier to catch mistakes and remove sensitive content, but the files still sit on Anthropic's infrastructure and are subject to the same legal and operational realities.

Why is Gemini's memory disabled in the EU?

Personal Context is excluded in the EEA, UK, and Switzerland because European data protection rules (primarily the GDPR) impose stricter requirements on automated profiling and default opt-ins. Rather than re-architect the feature, Google chose to restrict it by region. This is a recurring pattern in AI product launches.

What's the difference between Temporary Chat and a deleted chat?

Temporary Chat is designed to exclude the session from memory and training. Deleted chats are sessions the user explicitly removed. In practice, both can be subject to legal holds, backup retention, and safety review queues that keep data around far longer than the UI suggests. The Stanford 2025 study documented this gap in detail.

Can I move my memory between ChatGPT, Claude, and Gemini?

As of March 2026, yes, with caveats. Anthropic and Google both shipped Import Memory tools that pull context from competitor exports. ChatGPT has not yet reciprocated. Portability is improving, but round-trip fidelity (exporting, importing, exporting again) is still imperfect.

Should I turn off AI memory entirely?

For most users, the right answer is somewhere in the middle. Turn off implicit memory layers (chat history, Personal Context) if you value privacy and don't need heavy personalization. Keep explicit memory layers (Saved Memories, CLAUDE.md files) because you control them. And build a parallel user-owned knowledge base (like highlights or notes) that gives you AI context without giving a vendor your dossier.

How does Glasp fit into this?

Glasp is user-authored memory. Your highlights and annotations are text you wrote or selected, stored in a format you can export, delete, and carry between tools. When you use Glasp's AI chat, it reads from context you explicitly chose. This is the inverse of the default AI memory model, where vendors infer context from your behavior and store it opaquely.

Is AI memory regulated anywhere?

The EU's GDPR and the emerging AI Act impose indirect constraints, which is why Gemini's Personal Context is disabled there. In the US, state-level privacy laws (CCPA, CPRA, Illinois BIPA for biometric data) create partial coverage. There is no comprehensive federal AI memory regulation as of April 2026, though the NYT v. OpenAI case is influencing how courts treat deletion claims.

Conclusion: Your Memory, Your Legacy

The memory wars aren't really about AI. They're about who gets to write the authoritative record of your intellectual life.

For the first time, that record is being built by default, by systems you don't control, in formats you can't read, stored in places you can't fully audit. That's a significant change from the era when your knowledge lived in notebooks, bookmarks, and highlights you owned. The shift happened fast enough that most people missed it.

The good news: user-owned memory is still possible, and March 2026's portability moves make it easier to keep vendor memory as a convenience layer rather than a dependency. The choice now is less about which AI to use and more about where the canonical record of your thinking actually lives.

Your highlights, annotations, and curated context are a legacy. Vendor memory is a service. Be clear about which is which, because one will still be yours in ten years and the other will be whatever the current terms of service say it is.

Start owning your memory. Tools like Glasp exist for this reason, and the longer you wait, the more your digital self belongs to someone else.