ModelBrain Download (beta)
Articles

What leaves your computer when you use an AI memory tool

Matthew FirthBy , founder of ModelBrain · Oct 11, 2026 · 3 min read

Your text leaves your computer at two moments: when a memory is saved and when it is recalled. Recall always goes to the AI provider you are talking to, wherever the database lives. Save time is the only moment a memory tool can keep text from leaving, and that is where tools differ.

A small dark graphite cube with two thin cables leaving it, one ending in a single amber glow

What leaves, by tool and client

What leaves your computer at save time and at recall time, by memory tool and AI client. Checked October 11, 2026.
Tool and clientAt save timeAt recall timeSource
ModelBrain with Claude Desktop, Claude Code or CursorNothing leaves. The memory is saved on your computer and its meaning is calculated on your own processor.The assistant's provider receives what the assistant recalls.ModelBrain privacy page
ModelBrain with ChatGPTThe request passes through a ModelBrain relay in transit. The relay does not write the content to disk and keeps counts only.The same relay, then OpenAI receives what ChatGPT recalls.ChatGPT connection guide
ModelBrain sync and backup (paid, opt-in)The copy that is uploaded is encrypted on your computer first. The vault on your own disk is not encrypted. The service holds an account, an email address, device names, billing records, and the size and timing of uploads.Unchanged. Recall happens on your computer.Account guide and the privacy page
Mem0 open source, default settingsThe text is sent to the configured language model and embedder, which default to OpenAI's. The vector store is local by default in library mode.The assistant's provider receives what is recalled.Mem0 documentation

Mem0's overview lists OpenAI gpt-5-mini as the default language model and OpenAI text-embedding-3-small as the default embedding model, both needing an OpenAI key. It says you choose where each component runs, so a setup with no outside party at save time is possible, but it is not the default. The overview does not state where your data goes, so what happens beyond sending text to the configured models is for you to confirm. I did not check Mem0's hosted platform.

The save-time test

Recall is the same for every cloud assistant, so it cannot tell memory tools apart. Ask one question instead: what does the tool send when it saves a memory? Four checks answer it.

  1. Where is a new memory extracted? If a cloud model reads your conversation to decide what to save, that text leaves at save time.
  2. Where is its meaning calculated? Embeddings made by a hosted service send the text to that service. ModelBrain calculates them on your own processor.
  3. Where is it stored? A folder on your disk sends nothing. A vendor's servers do.
  4. What else is sent? Telemetry, crash reports, backups and sync. ModelBrain has no usage tracking, analytics or crash reports in the product.

What this does not cover

It says nothing about retention or how each company treats the text it receives. ModelBrain does not encrypt the vault on disk today, so turn on your computer's own disk encryption. Do not put the vault inside Dropbox, iCloud Drive, OneDrive or Google Drive, which would copy it to someone else's servers; ModelBrain checks for this at start and refuses to run. What your assistant does with a memory after receiving it is governed by that company's terms, not ModelBrain's. For the wider setup, see how to give Claude, ChatGPT and Cursor the same persistent memory. Last checked October 11, 2026; confirm any detail on the vendor's own page.

Common questions

Does a local AI memory keep my data private?

It keeps the stored copy on your computer. Any memory the assistant recalls is still passed to the AI provider you are talking to, so it leaves your computer at recall time.

What leaves my computer when a memory is saved?

It depends on the tool. With ModelBrain and Claude Desktop, Claude Code or Cursor, nothing leaves. With ChatGPT the request passes through a ModelBrain relay in transit, which does not write the content to disk. With Mem0 open source on its defaults, the text is sent to OpenAI's language model and embedder.

Does ModelBrain use my memories to train a model?

No. Its privacy page says memories are never used to train any model, and that there is no usage tracking or analytics in the product.

Sources