Claude Memory vs ChatGPT Memory vs Gemini Personal Intelligence

Claude Memory, ChatGPT's memory, and Gemini's Personal Intelligence all try to solve the same problem — not having to re-explain yourself every time you open an AI assistant — but they store different things, hand you different controls, and draw on different data. Anthropic's Claude saves work-oriented "topics" per project and keeps sensitive categories out by default; OpenAI's ChatGPT blends explicit saved memories with a broader reference to your chat history; and Google's Gemini Personal Intelligence, still in beta, reaches directly into Gmail, Photos, YouTube and Search rather than just your chat history. For most people weighing which one to trust with personal context, the short version is: Claude gives the most granular, project-scoped controls; ChatGPT has the widest reach across your chat history with a layered opt-out system; and Gemini has the broadest access to your real-world data but the narrowest rollout and default-off connections.

What "memory" actually means on each assistant

The three products are not the same feature wearing different names. Claude Memory is built around topics it saves while you chat — your role, the people you work with, your communication preferences, and ongoing project details — organized by project. ChatGPT's memory is really two features bundled together: "saved memories" (discrete facts it or you flag as worth keeping) and "reference chat history" (a broader, less itemized pull from past conversations). Gemini's Personal Intelligence is different in kind: rather than summarizing your conversations with the chatbot, it connects Gemini to your actual Google apps — Gmail, Google Photos, YouTube and Search — so it can retrieve a real email, photo or video when a prompt calls for it.

How Claude Memory works

According to Anthropic's own support documentation, Claude saves memory as individual topics while you chat, not as an end-of-conversation summary, and it can do this on its own or when you explicitly say "remember this." It covers your role and projects, the people in your work, your communication and technical preferences, and ongoing project details. Sensitive topics — health, politics, religion and similar categories — are not saved by default; you have to opt in under a dedicated setting. Some categories are excluded outright regardless of settings, including government ID numbers, criminal history, financial account numbers and immigration status.

Memory is scoped to projects: each project keeps its own memory, and inside a project Claude draws on that project's memory plus your general account memory, but never on another project's memory. You can isolate a project completely by turning off "Use account memory" on that project's page. For a one-off chat, you can turn memory off for just that conversation before you send the first message — the chat still saves to history and stays searchable, it just never feeds memory. Incognito chats go further: they are not saved to history or memory at all. Anthropic's product explainer covers the mechanics in more depth if you want the full walkthrough of how Claude's memory feature works.

How ChatGPT's memory works

OpenAI's help center describes memory as context ChatGPT can draw from several places at once: saved memories, custom instructions, files in your Library, and connected apps such as Gmail, depending on your plan and region. Crucially, "ChatGPT decides which available information is relevant to a response," and what it uses can change as your context changes — there is no fixed, user-authored list the way Claude's per-topic view works.

OpenAI splits the feature into two switches. Reference saved memories covers explicit facts ChatGPT stores separately from the chat that produced them, so deleting that original chat doesn't delete the memory. Reference chat history is a broader, continually-updated pull from your past conversations with no separate storage list; turning it off schedules that remembered information for deletion within 30 days, though the chats themselves remain in your history. Turning off saved memories also turns off chat-history reference, but not the other way around. OpenAI also runs an "improved memory" model alongside a legacy "saved memories" list in some accounts, and lets you toggle between them without losing chat history.

How Gemini's Personal Intelligence works

Google announced Personal Intelligence in a post by Josh Woodward, VP of Google Labs, Gemini app and AI Studio, describing it as a beta that lets the Gemini app combine reasoning with retrieval of specific details — an email, a photo, a video — once you connect your Google apps. In Woodward's own example, Gemini pulled a license plate number from a Google Photos image and cross-referenced a vehicle's trim level from a Gmail receipt to help him shop for tires.

The apps on offer at launch are Gmail, Google Photos, YouTube and Search, and you choose which ones to link — connecting is off by default. Google says Gemini doesn't train directly on your Gmail inbox or Photos library; it trains on a narrower set of data, such as specific prompts and responses, after filtering or obfuscating personal information. The company also built in guardrails around sensitive subjects: Gemini is designed to avoid proactively surfacing assumptions about things like your health, though it will discuss them if you bring them up. Google's own explainer covers the feature's connected-apps model in more detail if you're deciding whether to turn on Gemini Personal Intelligence.

Privacy posture, side by side

The three companies frame privacy differently because the underlying feature works differently. Anthropic leans on exclusion lists and project boundaries: certain categories (IDs, financial account numbers, immigration status) are never stored, sensitive topics require an opt-in, and a project can be fully walled off from your general account memory. OpenAI leans on layered toggles rather than exclusion lists — you control whether saved memories exist, whether chat history can be referenced at all, and separately whether your content trains the underlying model, with that last switch defaulting to off for Business, Enterprise, Edu and Healthcare workspaces. Google leans hardest on the opt-in framing because Personal Intelligence reaches outside the chat entirely, into Gmail and Photos: every app connection starts off, you choose each one individually, and Google states plainly that it does not train directly on your inbox or photo library, only on narrower, filtered prompt data.

How to actually delete what each one remembers

Claude offers two distinct levels: Pause, which keeps existing memories in place but stops Claude from using or creating new ones, and Reset, which permanently deletes all memory — including every project's memory — with no way to undo it. You view and edit individual topics under Settings, and you can also just tell Claude in chat to remember, change or forget something specific. Deleting a conversation does not delete the memories it generated; those have to be removed separately.

ChatGPT separates "memory" from "chat history" in a way that trips people up: deleting a saved memory stops it from being used going forward but doesn't retroactively scrub mentions of it from old conversations, and OpenAI's help center notes that fully removing a piece of information can mean deleting the saved memory, the chat where you first shared it, any related Library file, and any connected-app content where it also appears. The "Delete and turn off memory" option clears the memory summary and disables the feature, but OpenAI says it may retain logs of deleted memories for up to 30 days for safety and debugging.

Gemini's controls sit closer to account-level permissions than a memory list: you disconnect individual apps, delete your Gemini chat history, or turn Personal Intelligence off entirely, and Google says you can do any of that at any time. Because the feature retrieves from Gmail and Photos rather than storing a separate memory file, "deleting what Gemini remembers" largely means managing your underlying Google account data and your app connections rather than editing a bespoke memory summary.

Which plans and tiers actually get it

Claude's memory is the most broadly available of the three right now: Anthropic's own announcement and support documentation state it is on by default for Free, Pro and Max plans (mobile users need an up-to-date app), while Team and Enterprise require an organization owner to switch it on for members, and it's excluded entirely from organizations with HIPAA or public-sector agreements. Incognito chat, notably, is available on every plan regardless of memory status.

OpenAI is less specific in its own documentation, stating that "memory features and controls can vary by plan, region, platform, and workspace settings" and that a control's absence from your settings may simply mean it hasn't reached your account yet. The clearest carve-outs are for regulated use: improved memory is off by default for ChatGPT for Healthcare and for Enterprise accounts with a Regulated Workspace, and isn't covered under those workspaces' BAA, so protected health information shouldn't be entered regardless of the toggle.

Gemini's Personal Intelligence is the most restricted at launch: it rolled out as a beta to Google AI Pro and AI Ultra subscribers in the US on personal Google accounts, works on web, Android and iOS, and is explicitly unavailable to Workspace business, enterprise or education accounts. Google has said it plans to expand to more countries, to the free tier, and into AI Mode in Search, but as of launch it remains the narrowest of the three in both audience and feature maturity.

Which assistant should you trust with your personal context?

There's no single right answer, because the three tools are optimized for different kinds of "remembering." If you want an assistant that keeps separate contexts cleanly apart — one memory per client, per project, per team — Claude's project-scoped design and its explicit exclusion list for sensitive categories make it the most conservative default, and the Pause/Reset split gives you a clear, low-risk way to try it. If you want an assistant that already threads together everything you've ever told it across every chat, ChatGPT's reference-chat-history model goes the furthest, but that reach is also why its deletion story is the most involved: a single fact can live in several places at once, and clearing all of them takes more than one click. If you're comfortable letting an assistant reach into your actual inbox and photo library to save you from re-typing details you've already written down elsewhere, Gemini's Personal Intelligence is the most powerful in principle — but it's also opt-in app-by-app, US-only, subscription-gated, and still a beta that Google itself says can misfire with "over-personalization."

For anyone handling client or regulated work, the practical move is to treat all three the same way: read the sensitive-data defaults for the tool you already use, turn on only what you need, and use the private-mode option (Incognito, Temporary/Unpersonalized chats, or disconnecting an app) for anything you wouldn't want summarized later.

What none of the three have fully solved

Each company flags its own rough edges. Google is the most candid in its launch post, acknowledging that Personal Intelligence is a beta that hasn't eliminated mistakes, that "over-personalization" can connect unrelated topics, and that it can struggle with timing and nuance — Woodward's own example is Gemini wrongly assuming someone loves golf just because of golf-course photos. OpenAI's documentation acknowledges that its memory summary is a "non-exhaustive" view, meaning what you see under Settings may not reflect everything ChatGPT has actually retained, and that corrections like "don't mention this again" change future behavior without removing the original source. Anthropic's support article is comparatively quiet on failure modes, but is explicit that resetting memory is irreversible and that deleting a conversation doesn't automatically delete the memories it produced — an easy thing to miss if you assume clearing your chat history clears everything.

What next

If you're deciding where to start, the lowest-friction test is usually the private mode: open an Incognito chat in Claude, a Temporary chat in ChatGPT, or simply leave Gemini's app connections off, and see whether the assistant is still useful without memory before you decide how much personal context to hand over. From there, check each product's own settings page directly rather than relying on defaults, since all three companies describe availability as something that can vary by region, plan and even app version. And if the broader AI-assistant landscape matters to your decision — not just memory, but the assistants themselves — it's worth comparing how Claude Duet, OpenAI's Dots and Meta's Muse stack up for the kind of day-to-day, context-heavy work that memory features are ultimately built to support.

Written byTheo Park

Theo Park runs the AI desk at Pandromeda. He follows model launches from the frontier labs and the open-weight community, tracks the assistants and developer tools built on them, and explains what each release changes on pricing, capability and safety. His reporting leans on primary sources: model cards, technical reports, API documentation and the companies' own announcements.

More from Theo ParkAll AI coverage