Which AI has the best memory?

Every listicle answers this with a ranked list. None of them says what 'best' means. Best memory is four different questions — how much it holds at once, how long it keeps it, how many sources it spans, and whether you can read and correct what it kept — and the honest answer changes depending on which one you're asking.

The short answer

No one assistant wins on all four measures, and the differences between them are now smaller than the difference between all of them and what client-facing work actually needs. On the evidence of the vendors' own documentation, read on 2026-08-28:

• Claude has the most inspectable record. Anthropic: in Settings > Memory, "Everything Claude remembers is listed under Topics," each topic readable, editable or deletable individually — and when Claude leans on an earlier conversation, "you'll see citations linking back to the original chats."
• ChatGPT reaches the widest set of sources. OpenAI: memory draws on "your chats, files, and connected apps" — and a book icon under each answer shows which sources shaped it.
• Gemini gives the most explicit control over how long things are kept. Google publishes a default 18-month auto-delete on Gemini Apps Activity, changeable to 3 months, 36 months, or indefinite.
• None of the three remembers across sources you choose one at a time. Each remembers its own conversations, plus whatever whole accounts you connect to it. The thread you were CC'd on, the contract PDF, the voice memo after the call — none of that is in any of them unless you personally paste it in.

If you want the one-line version: Claude for a record you can audit, ChatGPT for reach, Gemini for retention control — and none of them for remembering the specific material your client work runs on.

Why the ranked lists can't answer this question

"Which AI has the best memory" gets answered with a top-ten because a top-ten is easy to write. But the four things people mean by "memory" pull in different directions, and a product that is excellent at one is often deliberately mediocre at another. A record you can read down line by line is necessarily a smaller record than a running synthesis. Retention you control tightly is retention that throws things away. Reach across your accounts is reach you had to grant wholesale.

So the useful move isn't a ranking. It's deciding which of the four questions is yours, and then reading the row that matters.

1. How much can it hold at once?

This is the question people usually think they're asking, and it's the one that matters least for remembering. The number involved is the context window — how much text the assistant can consider inside a single conversation. Anthropic's help center states that "Claude Opus 5 and Sonnet 5 support a 1M token context window on all paid plans when chatting with Claude," with 500K for Opus 4.8, Opus 4.7, Opus 4.6 and Sonnet 4.6, and 200K outside those models (verified 2026-08-28); the other vendors publish their own figures per model and plan, and they change often enough that any number quoted on a page like this ages within weeks.

The thing worth internalising is that a context window is not memory. It's the size of the desk, not the filing cabinet. Everything on the desk is swept off when the conversation ends. A million tokens is genuinely useful when you've pasted a 200-page contract in and want to interrogate it this afternoon; it does nothing for the question you'll ask in October about what a client said in March.

Which is why "which AI has the biggest context window" and "which AI has the best memory" are different questions with different answers — and why a shortlist built on the first one is answering the wrong thing.

2. How long does it keep it?

Here Google is the most explicit of the three, and deserves credit for it. In the Gemini Apps Privacy Hub, Google documents an auto-delete setting on Gemini Apps Activity with "the default of 18 months," changeable "to 3 months, 36 months, or indefinite." With Keep Activity off, chats are "retained with your account for 72 hours." And Google states plainly that chats reviewed by human reviewers "are not deleted when you delete your activity. Instead, they are retained for up to three years." You may not like every one of those numbers, but you are told what they are.

OpenAI frames the same question as a deletion job, and is candid about how wide it is. In OpenAI's own words: "To fully delete something ChatGPT may know about you, you'll need to delete every source where it appears, including past chats, archived chats, files, the memory summary, and disconnect any connected apps that may contain that information." Clearing one conversation, in other words, is not the same as clearing what was taken from it — a point OpenAI makes rather than hides.

Anthropic keeps the memory and the conversation apart, and defers the clock. Its help center: "When a conversation expires or is deleted, related memory entries generated from it won't be removed, but you can delete individual memories at any time." On duration it points elsewhere rather than publishing a window on the memory page: "All memory will be retained in accordance with existing chat data retention policies." Settings > Memory offers a pause and a reset, where reset "permanently deletes all memories including project memories."

The practical read: Gemini is best if you want a stated expiry you control. Claude gives you the cleanest per-item delete, but no expiry window is published on its memory page — with Claude, the forgetting is yours to do. Both are defensible; they're opposite answers to the same worry.

3. How many sources does it span?

On paper, ChatGPT reaches furthest. OpenAI describes memory as covering "your chats, files, and connected apps."

Claude's memory itself is built from your conversations — "Claude saves memory as a set of individual topics as you chat, rather than summarizing conversations after they end" — with a separate mechanism for the rest: connectors that Anthropic documents for "Gmail, Google Calendar, and Google Drive," "available for all users on Claude and Claude Desktop," which access your data "only ... when you explicitly ask a question or request an action requiring this information." That's fetching on request rather than remembering, and the distinction matters: it can answer from your mail today without that answer becoming part of what it knows tomorrow.

Gemini's reach is the most clearly documented and the most conditional. Google lists five connectable app groups — Contacts; Google Photos "(in certain countries and territories)"; "Google Workspace, which includes your content in services like Gmail, Calendar, and Drive"; Search services; and YouTube — with Gmail riding the single Workspace switch alongside Calendar and Drive. Two conditions catch people: Google states "Gemini Apps can't connect to Google Workspace when this setting is off," referring to Smart features in other Google products — a Workspace setting that Google's Gmail help documents changing from Gmail's own settings, where you "Under the General tab, scroll to Google Workspace smart features" and "Click Manage Workspace smart feature settings" (verified 2026-08-28); and the connected-apps feature is "not available in the European Economic Area, Nigeria, Switzerland, or the United Kingdom." If you're in Dublin or London, that row reads differently for you than for a reader in Chicago.

Notice what all three have in common. The unit of sharing is the account. You connect a mailbox, not a thread. There is no version of any of these where the answer to "what does it know" is "the four client threads I decided to hand it."

4. Can you see and correct what it kept?

This is the question that decides whether you can trust an answer in front of a client, and it separates the three most sharply.

Claude is ahead here, clearly. Anthropic: "Everything Claude remembers is listed under Topics." You "select any topic to read it, then use the edit icon to change it or select 'Delete' to remove it." The answers are traceable too: "When Claude references previous conversations, you'll see citations linking back to the original chats, along with the option to delete specific conversations," and where the Google Workspace connectors are in play, "Claude's response includes citations indicating which emails, calendar events, or documents were used as sources, with links to the originals when available." Projects are walled off from each other — "each project has its own separate memory space and dedicated project summary, so the context within each of your projects is focused, relevant, and separate from other projects or non-project chats" — which is a genuine advantage if you keep two clients apart, and a nuisance when the detail you want is in the other space. One plan boundary to know: Anthropic states that "searching past chats is available to users on paid plans (Pro, Max, Team, and Enterprise plans)."

ChatGPT gives you a summary and a source trail, but tells you the summary is partial. OpenAI's own words: the memory summary "will not include everything that ChatGPT remembers based on your chats," and, on whether it includes all your memories, "Not necessarily." Deleting cleanly is correspondingly involved — OpenAI: "To fully delete something ChatGPT may know about you, you'll need to delete every source where it appears, including past chats, archived chats, files, the memory summary, and disconnect any connected apps that may contain that information." Set against that, ChatGPT gathers the personalization sources in one place under the answer: "You can see what sources were used to personalize a response such as custom instructions, past chats, files, and memories by tapping the book icon below the response" — one control covering instructions, past chats, files and memories together, with the caveat, again OpenAI's own, that sources "may not show every factor or source that shaped a response."

Gemini has the least direct handle on the record. Its memory of past chats lives under Personal context and requires that you "be 18 or over," "sign in with a personal Google Account," and "have Keep Activity on"; Google notes it is "not available in certain features, like Gems or Live chats." To remove something it remembered, Google's route is to delete every chat containing that information from Gemini Apps Activity. And to find out whether an answer leaned on your history at all, Google's suggestion is to ask it: "Did you use any info from past chats?" That's a reasonable check. It isn't a list you can read down and prune.

The scorecard, criterion by criterion

Every quotation above was read on 2026-08-28 from the vendors' own help pages — OpenAI's Memory FAQ; Anthropic's "Use Claude's chat search and memory to build on previous context", "Use Google Workspace connectors", and "How large is the context window on paid Claude plans?"; Google's "Get personalization with memory of your past Gemini chats", "Connect your Google apps to personalize your Gemini experience", the Gemini Apps Privacy Hub, and the Gmail help page for smart features settings. All three products change these controls often and roll changes out to accounts at different times: your own settings screen is the authority, not this page.

Where each one is genuinely ahead

A comparison where one product wins every row isn't a comparison. Here's the fair version.

• ChatGPT — reach and explicability. It draws on the widest declared set of material, and the book icon under a response puts custom instructions, past chats, files and memories in one place, so what personalized the answer is visible in a single control. If you want one assistant that knows the most about you with the least effort, this is it.
• Claude — the auditable record. A topic list that Anthropic says holds everything it remembers, editable and deletable item by item, plus automatic separation between projects. If your instinct is "I want to read what it thinks it knows about me," Claude is built for you.
• Gemini — retention and Google reach. Documented auto-delete windows you choose, and the deepest reach into Gmail, Calendar and Drive if you live in Google and are outside the excluded regions.

And a fourth category is worth naming, because it shows up in these shortlists: the dedicated memory tools built for developers. Supermemory, for instance, describes itself as "the memory layer for AI agents. Context engineering platform powering enterprise APIs, developer plugins, and a personal app that remembers everything" (verified 2026-08-28 from supermemory.ai). If you're building software, that shelf is the right one — we go through it in Equerry vs Supermemory. If you're the person with four client relationships in flight, it isn't.

The question none of them is trying to answer

Run the four criteria back and something is missing from all of them. Every product above remembers its own conversations, plus whatever whole accounts you connect. Neither shape matches how a senior operator's material is actually arranged.

Consider a real week. A client writes on Tuesday to say the board review moved to the 14th and a new person now signs off on scope. On Thursday you dictate a voice memo walking to the car after the call where that got confirmed. On Friday a contract PDF arrives from their legal team. In October you need all three at once, and you need to be right in front of the client.

• None of that was typed into a chat window, so chat memory never saw it.
• Connecting your whole mailbox would reach the email — and also every other thread in it, which is a different decision than you meant to make.
• The voice memo and the photo of the whiteboard aren't in anyone's mailbox at all.

That's the gap. Not "which assistant remembers more," but "which one remembers the material I chose, from wherever it came, and can show me the original when I ask."

How Equerry answers it

Equerry is a personal assistant for iPhone, built around that question rather than around a chat log. The unit of sharing is the thing you decided to share:

• One thread at a time, not the mailbox. You get your own address — something like mark@mail.getequerry.app. CC it once on a live client thread and every later reply lands too, from whichever mail app you already read in. Nothing else in your inbox is reachable, because nothing else was sent.
• Everything else, the way you already capture it. Share a PDF or screenshot from any app, dictate a voice memo after the call, photograph the whiteboard, save a typed note. It's all searchable together, not filed in separate silos.
• Answers arrive with the original attached. Ask what the client said about the timeline and the answer carries a source chip — tap it and you're reading the email or note itself, before you repeat it on a call.
• A record you can read and prune. Every fact Equerry extracted is listed in Settings, each saved item shows what was taken from it, and anything you delete is gone along with what was derived from it.
• It's not built to expire. Nothing ages out on a schedule you didn't choose — what you saved in March is there in October, for the relationship, not just the conversation.
• Briefed before the call, not searching after it. When the meeting is on your calendar, Equerry can hand you a brief beforehand, assembled from what you've shared — including Thursday's late reply you never opened.

Two honest limits, because a page like this is worthless without them. Equerry doesn't sit in your meetings — it doesn't record or transcribe live calls; what it knows about a meeting is what you send it afterwards, whether that's your voice memo, your notes, or a transcript you forward. And it isn't a replacement for the assistants above at what they're good at: keep ChatGPT or Claude for the drafting and the thinking. Equerry is available on the App Store.

On where things rest, plainly, because it's two different answers. The content you create in the app — typed notes, voice memos, photos, PDFs — lives in your own iCloud account, on Apple's infrastructure, not in our database. A message you forward or CC arrives at our email provider, is read there and then deleted from it, and your own copy stays in your mailbox. But everything you share, from every one of those sources, produces the layer that makes it answerable: searchable excerpts, embeddings, and extracted facts. Those derivatives do sit on our servers — that is true of the voice memo and the photo just as much as of the forwarded email — and every extracted fact is listed in Settings → Your Memory, deletable on a tap. The full account is in our privacy policy, and it's worth ten minutes before any client material goes into any tool, ours included.

So which should you choose?

• You want the record readable and correctable, item by item → Claude.
• You want one assistant that quietly knows the most, with source hints under each answer → ChatGPT.
• You want stated retention windows you control, and you live inside Google → Gemini (check the regional exclusion on connected apps first).
• You want a big window for a document you're working through today → any of them; this is the context-window question, not the memory one.
• You want an assistant that remembers the specific emails, PDFs, voice memos and notes you handed it — and shows you the original → that's a different shape of product, and it's the one Equerry was built to be.

The uncomfortable conclusion for a page that's supposed to crown a winner: for most people asking this question, the answer isn't a swap. It's that the assistant you already pay for remembers its own conversations perfectly well, and the thing you're actually missing was never in a conversation.

Last updated 2026-08-28.

All product names, logos, and brands are property of their respective owners. Equerry is not affiliated with, endorsed by, or sponsored by any other product mentioned; comparisons are provided for informational purposes.