How AI Memory Works in 2026 (ChatGPT, Claude, Gemini & More)

TL;DR

Six assistants, one question: does the AI remember you, and can you make it stop? This page audits official documentation for ChatGPT, Claude, Gemini, Perplexity, Microsoft Copilot, and Meta AI, current as of August 1, 2026. Only Google Gemini confirms memory is on by default. Four of six let you delete a single memory entry, Gemini is all-or-nothing, and Copilot can bring deleted memories back within 30 days without telling you. Only Perplexity states outright that memory content trains its models by default. We also cover the 2026 "Memory Heist" exploit against Claude, a court order compelling OpenAI to hand over 20 million ChatGPT logs, and the overturned Italian Garante fine, plus exact click paths to view, edit, and delete your memory in each assistant.

I spent a week last month asking six different AI assistants to explain their own memory settings, and got six different answers with six different levels of honesty.

Some told me exactly what they store. One made me dig through a help center with no dedicated memory article at all. That distance between assistants is the whole story here.

Last verified: August 1, 2026. Every claim below traces back to an official help-center article, product announcement, or a named, dated report. Where the docs themselves are silent, we say so instead of guessing.

What's Inside

Does ChatGPT remember everything I tell it, or only what I ask it to save? Is any assistant's memory on without asking me first? Which ones let me delete a single memory instead of wiping everything? Does deleted memory stay deleted? Is my memory used to train the model, and can I turn that off? What happened with Claude's "Memory Heist," and does it affect me? And how much of my memory can I take with me if I switch tools?

The Six-Assistant Memory Matrix

I read the primary documentation for all six assistants start to finish, not just the landing pages. The clearest finding: only Google Gemini states, in its own words, that memory is on by default. Everyone else leaves the default undocumented or splits the answer by account type.

(Full sourcing for every cell: OpenAI Memory FAQ, Claude's chat search and memory, Gemini personalization launch, Perplexity Help Center, Microsoft 365 Copilot memory settings, and PCMag's reporting on Meta AI.)

A few numbers from that table are worth sitting with on their own.

Only one of six assistants documents memory as on by default. Google states it plainly: "This setting is on by default to help Gemini give you more relevant responses" (Google Blog). Claude splits the difference, off for individuals, on for Enterprise orgs. The other four don't say what happens the first time you log in.

Four of six let you delete a single memory without nuking the rest. ChatGPT, Claude, Perplexity, and Meta AI support deleting one entry at a time. Gemini doesn't give you that button. Its own help page tells you to "delete all chats with this info from Gemini Apps Activity" to remove a single remembered fact (Google Support). That's a chat delete with a side effect, not a memory delete.

Copilot's story is the strangest one I found. Microsoft's own documentation adds a catch that's easy to miss: "If the Chat history personalization setting is turned back on within 30 days after the deletion, Copilot adds back the deleted memories from your chat history" (Microsoft Support). Delete something, flip a setting back a week later, and it reappears.

Perplexity is the only one of the six that states, in plain language, that memory content feeds model training by default: "We may use content from memory to improve our models for everyone" (Perplexity Help Center). ChatGPT, Claude, Gemini, and Copilot avoid a direct default statement in the docs checked. Claude and Gemini both document a memory import feature, letting you bring context from other AI apps into a new one. We'll come back to why that matters more than it sounds.

Assistant-by-Assistant Control Walkthroughs

ChatGPT: two memory layers, one settings screen

ChatGPT runs two systems side by side. "Saved memories" are things you told it directly, or that it decided on its own were worth keeping: "If you share information that might be useful for future conversations, ChatGPT may save those details as a memory without you needing to ask" (OpenAI Memory FAQ). The newer "memory summary" layer is fully inferred.

To control it, go to Settings, then Personalization, then Memory. View the summary, tap the book icon under a response to see which memories informed it, delete individual entries, clear everything, or turn memory off. One catch: removing a saved memory "won't remove mentions from past chats." Delete the source conversation too if you want it fully gone.

Claude: memory that lives inside your projects

Claude organizes memory into categories: your role, working style, coding preferences, and ongoing project details, updated "in real time as you chat rather than on a fixed daily schedule" (Claude Support). It's off by default for individual accounts. Turn it on in Settings, then Memory.

Once on, the Memory panel lists everything, grouped by category, and you can delete any single entry. Two disable modes matter here: Pause keeps existing memory but stops building new memories, while Reset "permanently deletes all memories including project memories" and cannot be undone. Enterprise orgs get the opposite default, memory generation is on org-wide unless an admin turns it off.

Gemini: personalization you mostly have to talk to

Gemini's memory draws on three sources: past chat context, connected Google app activity, and explicit "Saved Info" instructions (Google Support). This is the one assistant Google confirms is on out of the box.

There's no dedicated list of inferred memories to scroll through. The official guidance is to ask directly: "To check if Gemini used your past chats, just ask Gemini, 'Did you use any info from past chats?'" To delete something, delete the chats that contain it, and if a connected app is also a source, disconnect that too. Two steps, not one.

Perplexity: memory, search history, and Brain

Perplexity separates three layers: Memories (frequency-weighted, built from what you ask about most), Search history (your literal past queries), and Brain, a self-improving system on Max and Computer plans (Perplexity Help Center). "Memory makes Search personal. Brain gives Computer working context."

Manage Memories opens your full memory bank, searchable and filterable, with a trash icon on every entry. Referenced memories show up inline in your answer's sources, the same way a web citation would. Memory is scoped to individual projects by default. Incognito mode exempts you from all of it, nothing asked there is retained or used in memory. Brain is the layer I have the most open questions about, and I published the 20 questions I'm still asking about it as I test it against my own workflows.

Microsoft Copilot: memory plus an IT admin in the loop

Microsoft 365 Copilot separates Copilot Memory (explicit, auto-saved facts) from chat-history personalization (inferred patterns) (Microsoft Support). Deleting fully means two actions: delete the saved memory in Chat settings, then turn off chat-history personalization, or ask IT. Remember the 30-day auto-restore quirk from the matrix above before assuming a deletion is final.

Temporary Chat mode exists too, and won't show up in your history. It's still subject to your organization's retention policy behind the scenes, since this is fundamentally an enterprise product with an admin above the user settings.

Meta AI: memory that only activates when you ask

Meta AI's memory only builds inside 1:1 chats with the assistant, not passively across WhatsApp or Messenger, and excludes group chats. Per Meta's statement to press, it will "only remember certain things you tell it in 1:1 conversations, and you can delete its memories at any time" (PCMag).

Tap the assistant's name to open its Memory view, where every save appears as a "Memory Updated" prompt. Delete individual entries by tapping and holding, or wipe everything with the /reset-ai command. Worth flagging honestly: Meta doesn't publish one consolidated memory help article the way the other five do. We pieced this together from a press statement and general WhatsApp FAQ pages, itself a data point about how seriously each vendor documents this feature.

The Builder's Angle: Memory as Switching Cost

Independent technical essayist Shlok Khemani argues OpenAI built ChatGPT's memory system the way it did specifically because "memory also creates lock-in" for a super-assistant used across search, coding, and daily tasks (shloked.com). Industry analysis from RiffOn goes further, calling accumulated memory "the only real moat in AI" in a market where switching between frontier models costs almost nothing (RiffOn).

That thesis has a hole in it, and the hole is documented. Claude ships a memory import feature that lets you bring context in from other AI assistants. Gemini ships one too: "To bring your memory from other AI apps to Gemini, just head to your Settings and select the new import option" (Google Blog). Vendors are building the moat and handing out the ladder at the same time. Memory is also only one layer of what an assistant carries into a task. How you feed the rest of that context is a builder skill in its own right, one I unpacked in the context engineering guide for product builders.

We don't have a clean, company-disclosed study proving memory features drive retention. What we have is a widely argued thesis from industry commentators, and two vendors undercutting their own advantage without seeming to notice.

2026 Incidents Worth Knowing About

The Claude "Memory Heist." In mid-2026, security researcher Ayush Paul showed that Claude's web_fetch tool could be tricked, through a disguised website, into leaking memory-stored personal details one character at a time, encoded into a sequence of clicked links. The exfiltrated data included Claude's own inferred guess at the user's hometown, deduced from an offhand mention of a hackathon. Paul reported it through Anthropic's HackerOne program, and Anthropic disabled the tool's ability to follow links within fetched pages, closing the path (GIGAZINE). The vulnerability wasn't in how memory got stored. It was in how an agentic tool could move already-stored memory out the door without the user noticing.

20 million ChatGPT logs, compelled by court order. In January 2026, a US federal judge affirmed an order forcing OpenAI to produce 20 million anonymized ChatGPT conversation logs in ongoing copyright litigation (Jones Walker LLP). This is a copyright case, not a memory case. It still matters, because it proves the same chat logs feeding both "saved memories" and "reference chat history" can be pulled into legal discovery at scale, regardless of what you've personally deleted.

The overturned Italian fine. On March 18, 2026, Italy's Court of Rome annulled the Garante's €15 million fine against OpenAI over ChatGPT's handling of personal data, wiping out what had been the only final GDPR enforcement action in Europe tied to the launch of generative AI (Cross-Border Data Forum). That case concerned training data broadly, not the memory feature. No EU ruling has directly addressed whether a "right to erasure" fully applies to an AI's inferred, synthesized memories. That question is open, not settled.

FAQ

Does ChatGPT remember everything I say to it?

No. ChatGPT keeps two separate systems: explicit "saved memories" you or it flagged as worth keeping, and a newer "memory summary" that updates automatically based on what it judges important. I dug through OpenAI's own FAQ and the default on/off state for new accounts isn't stated anywhere in it, which is a blindspot worth knowing before you assume it's tracking every word.

Is AI memory on by default?

Only Google Gemini says so outright, in its own documentation: "This setting is on by default." Claude splits the answer by account type, off for individuals, on for Enterprise orgs. ChatGPT, Perplexity, Copilot, and Meta AI leave their default undocumented. I'd love to tell you all six, but the receipts only support one.

How do I delete AI memory?

I read the deletion mechanics for all six instead of trusting the settings label alone. Four of six, ChatGPT, Claude, Perplexity, and Meta AI, let you delete a single memory entry directly. Gemini requires deleting the entire source chat instead of one fact. Copilot needs two separate steps and can auto-restore a deleted memory within 30 days if you flip chat-history personalization back on. Check the walkthrough above before assuming "delete" means what you think it means.

Does AI memory get used to train the model?

Sometimes, and often you won't find out from the settings page. Perplexity is the only one of the six that states this plainly in its help center: memory content may be used "to improve our models for everyone" by default, with an opt-out toggle. ChatGPT, Claude, Gemini, and Copilot don't give a direct default answer in their documentation. Claude at least frames it as a choice you make, rather than a silent default.

Can I move my AI memory from one assistant to another?

Yes, with two of six. Claude and Gemini both document memory import features, letting you bring context in from other AI apps. That's a real crack in the idea that switching assistants means starting over from nothing.

Audit your own memory settings once a quarter, test "delete" claims yourself instead of trusting the settings label, and don't assume training defaults without checking the doc. We'll keep this page updated as vendors change theirs. If you catch a discrepancy between what's written here and what you see in your own account, tell us 🤗