Articles

Can ChatGPT use a memory that lives on your computer?

By Matthew Firth, founder of ModelBrain · Oct 22, 2026 · 3 min read

Yes. The ChatGPT desktop app can run an MCP memory server on your computer, and ChatGPT on the web can reach one only through a hop that is reachable from the internet. Whether ChatGPT can also save to it depends on your plan. Here is our own proposal, the hop-and-write test: for any route, ask who runs the hop and whether ChatGPT can write to your memory or only read it.

A laptop outline holding a glowing amber capsule, one amber line running through a small hexagonal relay to a cloud outline and a second shorter line running to a small square window

Three routes, scored on October 22, 2026

How ChatGPT can reach a memory on your computer. Rows one and two are from OpenAI's pages, row three from ModelBrain's docs.
Route	Who runs the hop	Can ChatGPT save to it?	Where this comes from
ChatGPT desktop app, local STDIO server	None. The server runs as a process on your computer	OpenAI's page lists no plan limits for this route	OpenAI's MCP page
ChatGPT on the web, Secure MCP Tunnel	You run OpenAI's `tunnel-client`. OpenAI hosts the tunnel endpoint it connects out to	Business and Enterprise/Edu: yes, with confirmation. Pro: read and fetch only	OpenAI's help page and tunnel guide
ChatGPT on the web, hosted relay (ModelBrain)	ModelBrain runs the relay. It does not write request or answer content to disk	We have not confirmed saving on each ChatGPT plan	ModelBrain docs

OpenAI's help page says ChatGPT on the web cannot connect to localhost directly. It connects to remote MCP servers, and a server on a private network or a developer machine needs a Secure MCP Tunnel. The page also says custom MCP connections are web only and not on mobile.

What the tunnel hop is

Per OpenAI's guide, a tunnel is an outbound-only connection from a host inside your network to an OpenAI-hosted endpoint. You run the client and the memory server, and OpenAI runs the endpoint. Setup needs a Platform organization linked to your ChatGPT workspace. The guide gives Enterprise and Edu examples and does not say whether personal ChatGPT plans can use it. It also does not state how long OpenAI keeps transport payloads or queue records.

What the relay hop is

ModelBrain's route is a plugin you add in ChatGPT's settings, pointing at a relay we run. Per our ChatGPT docs, the relay ties your ChatGPT session to the computer you picked, forwards each request to it and the answer back over an encrypted connection, and holds only the connection state needed to keep the login alive. You need a free ModelBrain account. Claude Desktop, Claude Code and Cursor do not use the relay, because they talk to your local ModelBrain directly.

The write question

A memory ChatGPT can only read is a lookup, not a memory. OpenAI's help page says Pro users can connect custom MCP servers for read and fetch actions only, while Business and Enterprise/Edu connections can include write actions. We have not confirmed how that applies to a plugin set up the way our docs describe, on each ChatGPT plan. Check your own plan by asking ChatGPT to save one fact and then recalling it from a new chat or another assistant.

What stays true on every route

A recalled memory is sent to ChatGPT's model, so files on your computer do not keep the whole exchange local. See what leaves your computer and three readers to check.

Sources

Connecting custom MCP servers in ChatGPT, OpenAI Help Center
Secure MCP Tunnel, OpenAI developer docs
MCP servers, ChatGPT docs
ModelBrain and ChatGPT, ModelBrain docs

Common questions

Can ChatGPT connect to localhost? Not from the web. OpenAI's help page says ChatGPT cannot connect to localhost directly and needs a remote server or a Secure MCP Tunnel. The desktop app can run local MCP servers as processes on your computer.

Can ChatGPT write to a local memory on a Pro plan? OpenAI's help page says Pro users can connect custom MCP servers for read and fetch actions only, and that Business and Enterprise/Edu connections can include write actions. Plans change, so confirm in your own account.

Does ModelBrain work with ChatGPT? Yes, on the web, through a relay we run that is free with a ModelBrain account. The relay does not write request or answer content to disk. Claude Desktop, Claude Code and Cursor connect to ModelBrain directly instead.

Last checked October 22, 2026. Settings and tool behavior change often, so confirm a detail on the vendor's own page before relying on it.