Security

Last updated: 8 September 2026

The goal of this page is narrow: to let you verify our claims from the architecture itself, rather than take our word for them.

The short version

In local-only mode, Cognitive Atoms runs on your device and stores its core records locally. Your captured thinking is not held on our servers, so there is no copy of it for us to read, lose, subpoena, or sell. That is a property of how the system is built, not a promise about how we behave.

With a hosted account (beta) — including the Claude connector — your captured thinking is stored server-side, in your own account, scoped to your account at the database layer, and used only to serve your own capture, recall, and continuity. That copy exists so your thinking can follow you across devices; it is never sold or used to train AI models, and you can delete it yourself, along with the whole account, as described in the Privacy Policy.

Where your cognition lives

In local-only mode, on your machine. The runtime processes and stores your records locally, and no cloud account is required for it to work. There is no cloud relay for your captured thinking, no background sync to us, no telemetry, no advertising profile, and no sale of your data. With a hosted account (beta), your cognition lives server-side in your own account — see §4.5 of the Privacy Policy — and there is still no telemetry, no advertising profile, and no sale of your data.

In local-only mode, because we never hold your cognition, deletion is not a request you have to trust us to honour: remove the Cognitive Atoms data on your device and it is gone — we have no copy to delete. For a hosted account we do hold a copy, so deletion is a request we have to honour — and you start it yourself: Profile in the apps, or your profile page on the web, deletes the whole account on a typed confirmation. The bare account record is removed by us afterwards and confirmed within 30 days. The Support page and §4.5 of the Privacy Policy list exactly what goes, and the three things that do not.

How capture actually works

Nothing enters the system on its own. Every record follows the same path:

You choose

A conversation, a selected passage, or a note. Cognitive Atoms does not monitor your screen, record meetings, or scrape your workspace.

You review

You see what was captured before it is saved. Nothing is written to your record without passing in front of you first.

It is stored in your record

The record is written to your device in local-only mode, or to your own hosted account (beta) — and nowhere else.

You carry it forward

When you choose to work with ChatGPT, Claude, Gemini, or another provider, Cognitive Atoms prepares the relevant context. It does not send messages or take actions for you.

Fidelity: nothing invented

What Cognitive Atoms shows you back is quoted from what you brought it. It does not summarise your thinking into a claim you did not make, score you, or infer traits about you and present them as fact.

This is a deliberate limit, not a missing feature. A system that invents confident conclusions about how you think is not preserving your reasoning — it is replacing it with its own. Where we show you something, you should be able to trace it back to the words you captured.

Provider independence

AI providers are tools you choose, not owners of your memory. You connect your own provider account or key — we do not resell provider access, and we never fall back to a key of ours. In local-only mode that credential stays on your device and we never receive it. With a hosted account you give it to our server, which is what calls the provider for you; it is stored encrypted at rest in a slot belonging to your account, and one account’s credential is never used to serve another’s request.

This is also what makes the continuity claim real rather than rhetorical: because your record is yours and provider-independent — on your machine, or in a hosted account scoped to you at the database layer — switching from one model to another does not reset you to zero.

What this does not protect against

Being straight about the boundaries matters more than a longer list of assurances.

Your device is the trust boundary. If your machine is compromised, or another person has access to your unlocked device, your local records are exposed. Local-first moves the risk to you — it does not remove it.
What you send to a provider is governed by that provider. Once you carry context into ChatGPT, Claude, or Gemini, their terms and retention policies apply to that content. We cannot extend our guarantees into someone else's system.
A hosted account puts us inside your trust boundary. The cognition in a hosted account sits in our database, so it is something we could in principle read, lose, or be compelled to hand over. Scoping every row to your account stops other users reaching it; it does not stop us. If that is not a trade you want to make, local-only mode leaves us holding nothing of yours.
Your waitlist email is held by us. It is unrelated to your cognition. Write to privacy@cognitiveatoms.com to remove it.
This is early software. The beta is evolving, and we would rather tell you that than imply a maturity it has not earned yet.

Reporting a security issue

Please write to security@cognitiveatoms.com rather than disclosing publicly. We will work with you, and we will not pursue good-faith research.

It helps to include what you were doing, what you expected, and what happened instead. Please do not send us the contents of your captured cognition — we do not need it, and we would rather not have it.

The principle underneath all of it: your thinking should remain yours — portable, inspectable, and useful over time. Where we can make that structural rather than promissory, we do.