How ChatGPT, Claude, and Gemini's built-in memory works today — and what changes when memory follows you between them.
By Anthony Conti · Astra AI, LLC ·
Persistent memory, for an AI assistant, means what it knows about you is stored outside any single conversation and survives closing the chat, restarting the app, or switching to an entirely different model. Every major assistant now has some version of this inside its own walls. What none of them do natively is let that memory follow you when you leave — and for most people who use more than one AI tool in a week, that is the part that actually matters.
The test, in one sentence.
Open a different AI assistant than the one you used yesterday and ask it something only yesterday's conversation would explain. If it has no idea, your memory is trapped in one product — persistent within it, but not actually persistent across the tools you use in a normal week.
How built-in memory works today
Verified against each vendor's own documentation this session — this changes quickly, so treat this as a snapshot rather than a permanent spec.
Product
How it remembers
Leaves the app?
Cost
ChatGPT Memory
Saved memories (an editable list of facts) plus implicit recall from past chats
No — memory is visible and usable only inside ChatGPT
Included: basic memory on Free; advanced memory on Plus (~$20/mo) and Pro (~$200/mo)
Claude memory
An automatically extracted summary of preferences, projects, and key facts, organized into browsable "Topics" (reorganized August 2026); unified across chat and Cowork
No — it is a Claude-specific retention layer, separate from Projects
Included with a Claude account
Gemini Memory
A "Personal Intelligence" layer that learns continuously from chat activity, alongside explicit "Saved info"
No — and it requires Gemini Apps Activity to be on; unavailable in the EEA, UK, Switzerland, and Nigeria
Included with a personal Google account, 18+
MIND (MCP memory server)
A knowledge graph any MCP-compatible client can read and write — Claude, ChatGPT, Gemini, Cursor, and more
Yes — that is the entire point
Free tier available; see /mcp-memory-server/setup
Claude's memory does not store a word-for-word transcript — it stores an automatically extracted summary, reorganized around browsable "Topics" in an August 2026 update that unified memory across chat and Cowork. Gemini's Memory requires Gemini Apps Activity to be enabled and is unavailable to users under 18 or in the EEA, UK, Switzerland, and Nigeria.
Why single-model memory has a shelf life (and what carries across models)
Every product in the table above solved the persistence problem inside its own walls, which is real progress over 2023's completely stateless chatbots. What it does not solve is portability: preferences saved in ChatGPT do not exist in Claude, and a fact Claude extracted about your project does not exist in Gemini. The practical effect is that switching models — trying a new release, using whichever tool a client mandates, or simply preferring a different assistant for different tasks — means re-teaching your new AI everything the old one already knew.
This is not a hypothetical inconvenience for most people who work with AI daily in 2026. A single team routinely uses Claude for one kind of task, ChatGPT for another, and Gemini for whatever is already wired into their Google Workspace — plus Cursor or Windsurf for anything in an editor. Every one of those tools maintaining a separate, walled-off memory means the same preferences, the same project context, and the same corrections get re-explained in up to four different places, indefinitely, with no single one of them ever catching up to what the others already know.
The cross-model wedge: MCP as the portable layer
The Model Context Protocol, the open standard behind MCP memory servers, exists precisely to make an integration like this vendor-neutral. Instead of memory living inside one product, it lives in a server any MCP-compatible client can call — which today includes Claude, ChatGPT, Cursor, Windsurf, and a fast-growing list of others. Practically, this means the same underlying memory — the same agent memory — is available whichever assistant you open next, rather than each assistant maintaining its own separate, non-transferable copy. This is not a hypothetical: MCP is already the connection method ChatGPT, Claude, Cursor, and Windsurf each support today, which is the specific reason it works as a cross-model layer rather than a one-vendor integration.
How to set up memory that follows you
The mechanics are the same MCP connection covered in full at the setup guide: one paste for Claude Code (claude mcp add --transport http mind https://www.m-i-n-d.ai/mcp), one custom-connector URL for Claude Desktop and claude.ai, and the same connector URL in a small JSON file for Cursor and Windsurf. Once connected from more than one tool, all of them are reading and writing the same underlying memory — which is the entire difference between this and a built-in memory feature.
What should you check before trusting any memory product with something real?
Persistent memory is, definitionally, a system that keeps your information around longer than a single conversation — which makes the trust questions worth taking seriously before you wire it into your daily workflow, MIND included. None of the five checks below are exotic; they are the same baseline anyone would apply to a password manager or a cloud backup, and a memory product asking to hold real personal or business context deserves the same scrutiny.
Can you see exactly what was stored? A memory feature you cannot browse is a black box — you have no way to catch it storing something wrong, or something you would rather it had not kept.
Can you delete a specific memory, not just clear everything? ChatGPT, Claude, and Gemini all support per-item editing today (see the table above) — a memory system without granular deletion is behind the baseline the major vendors already ship.
Is every connected tool visible and individually revocable? When memory is reachable by multiple AI clients through MCP, you should be able to see every tool that has ever connected and cut off any one of them without affecting the others — check this in the product's own settings before relying on it.
Does the vendor say what it does with your data? Whether memory is used for anything beyond serving it back to you (model training, analytics, sharing) should be stated plainly, not buried — this is true whether the vendor is Anthropic, Google, OpenAI, or Astra AI.
What happens if you stop paying, or the company shuts down? A proprietary memory format with no export path means your accumulated context has an expiration date tied to a business decision you do not control. Portable, standards-based storage (MCP included) at least keeps the door open.
Isn't ChatGPT's memory or Claude's memory already persistent?
Within that one product, yes — genuinely so. The limitation is portability, not persistence: ChatGPT's saved memories don't travel to Claude, and Claude's memory doesn't travel to Gemini. If you only ever use one assistant, built-in memory may be all you need. The moment you move between models — which most heavy AI users now do routinely — that memory has to be rebuilt from scratch on the new one.
How is this different from a second brain?
A second brain is the broader, consumer-facing idea — an external system that holds what you know. Persistent memory, as used on this page, is specifically the part of that system an AI reads and writes automatically, without you filing anything by hand. Our /second-brain page covers the wider concept.
Does cross-model memory require MCP specifically?
Not in principle — you could build a proprietary sync layer between two specific products. In practice, MCP is the standard every major AI client (Claude, ChatGPT, Cursor, Windsurf, and more) has converged on for exactly this kind of integration, which is why an MCP memory server is the practical way to get it today rather than waiting for vendors to build point-to-point integrations with each other.
Is persistent memory private?
It should be scoped and revocable regardless of which system you use — every connected tool should be visible and individually disconnectable from a settings page, and you should be able to see exactly what's stored. Check this before trusting any memory product with anything sensitive, MIND included.
MIND is an MCP memory server backed by a real knowledge graph — the memory layer Claude, GPT, Gemini, Cursor and Windsurf all read from. Free tier, no credit card.