Last updated: 8 September 2026
Cognitive Atoms runs in two ways: local-only, where your thinking stays on your own machine, and a hosted account (in beta), where the cognition you capture is stored server-side in an account that belongs to you. This policy describes both honestly — the data we handle in each mode, and the things we have built ourselves so that we cannot handle.
Cognitive Atoms, Inc. (“Cognitive Atoms”, “we”, “us”) provides a local-first tool for preserving and working with your own thinking.
For privacy questions, or to exercise any right described here, contact privacy@cognitiveatoms.com.
If you use Cognitive Atoms from inside an AI app such as ChatGPT, that app has to reach the Cognitive Atoms instance running on your computer. Your computer is not directly reachable from the internet, so the connection passes through a relay service we operate. This section describes exactly what that relay does, because it is the one place your data moves through infrastructure of ours.
This section describes the local-only path. The hosted Claude connector works differently — it talks to our hosted service, where your account’s cognition is stored, and is described in §4.5.
The relay cannot read what passes through it. The encrypted session is established directly between the AI app and your own machine. Your machine holds the private key; the relay does not have it and never receives it. The relay forwards encrypted bytes it has no ability to decrypt.
ywzccc…42a.relay.cognitiveatoms.com. It is pseudonymous: it is not
your name, your email, or your account, and it is not derived from any of them.Routing labels exist in the relay’s memory only for as long as a connection is active, and are discarded when it ends. They are not written to disk and not logged. We do not build a history of which machines connected, or when.
Because we hold no key to your traffic and no record linking a label to a person, there is no store of your cognition for us to search, disclose, or hand over — including in response to a legal demand. That is a property of how the system is built, not a promise about how we behave.
When you join the waitlist we store the email address you submit and a truncated user-agent string. We use them to contact you about the beta and to diagnose signup problems. The form also contains a hidden anti-spam field, which is discarded.
Everything you capture is written to a store on your own device. We have no copy. Uninstalling the application, or deleting that store, removes it.
What you capture in the browser is sent only to the Cognitive Atoms runtime on your own machine.
If you connect an AI provider in local-only mode, requests go from your device to that provider under your own account or key, and the key or OAuth token you connected stays on your machine — we never receive it. Their handling of that data is governed by their policy, not ours. If no provider is connected, AI features fail rather than silently falling back to a key of ours.
With a hosted account, the provider key or OAuth token you connect is sent to our server, because our server is what calls the provider for you. We store it encrypted at rest, in a slot belonging to your account, and one account’s provider credential is never used to serve another account’s request. We still do not resell provider access and still never fall back to a key of ours. Deleting your account deletes the credential (§4.5).
You can also use Cognitive Atoms with a hosted account: you sign in with an email address and password, and the cognition you capture through hosted surfaces — including the Cognitive Atoms connector for Claude — is stored on our servers, in your account, so your thinking can follow you across sessions and devices. This means we hold a copy of what you capture. That is the trade a hosted account makes: the local-only mode described in §4.2 sends us nothing, and this one does.
The connector receives only the explicit arguments of a tool call — the text you ask it to capture, or the question you ask it to recall against. It does not receive your conversations beyond that, and it never reads Claude’s memory, chat history, summaries, or files.
Hosted accounts run on infrastructure we operate (hosted on Railway, with the database on Supabase — see §5). Every record is scoped to your account, and that scoping is enforced at the database layer (row-level security), not just in application code.
Your hosted cognition is used for exactly one thing: serving your own capture, recall, and continuity. It is never sold, never used for advertising, never used to train AI models, and never shared with other users. Nothing leaves your account except through your own explicit action, and any sharing you initiate passes through the Sharing Gate — our screening layer for outbound projections.
You can disconnect an AI app at any time — from that app’s own settings or from your Cognitive Atoms account — and all tokens for that grant are invalidated.
You can delete your whole account from inside the product — from
Profile in the iPhone and Android apps, or from your account’s profile page on the
web. You confirm by typing the phrase delete my account, and it cannot be
undone.
Confirming it removes, immediately: everything you have captured in the hosted account, and the provenance rows recording where imports came from; every source you connected (calendar, Google Workspace, Outlook, Slack, YouTube) and the tokens held for it; the AI-provider keys and OAuth connections you added; every AI-app connection grant, with its whole access- and refresh-token family; your connector keys, the sessions delegated from them, and their audit trail; your beta cohort membership, and any invite you redeemed; and your projects and record overlays. If a step fails, the app tells you so rather than reporting success.
Three things it does not remove. We would rather name them than imply a cleaner sweep than the code performs:
You can also email privacy@cognitiveatoms.com from your account’s email address and we will do all of it for you within the same 30 days.
The mobile apps let you dictate instead of type. The microphone is on only while you are dictating, and it stops the moment you stop. No audio file is recorded, kept, or uploaded — not to your device’s storage, and not to us. What the microphone produces is turned into text, and that text goes straight into the note or message you are writing, where you can edit it or discard it like anything you typed.
Recognition happens on your phone. On Android, Cognitive Atoms asks the system for on-device speech recognition and uses nothing else; if your phone cannot do it on the device — because it is an older Android release, or has no speech model for your language — dictation tells you so and declines, rather than sending your voice to a server. On iPhone, Cognitive Atoms asks Apple’s Speech framework for on-device recognition wherever your language supports it; for a language Apple cannot yet recognise on the device, Apple performs that recognition on Apple’s own servers under Apple’s privacy policy, and iOS asks your permission before any of this begins.
Nothing about dictation reaches our servers. Your speech is never sent to Cognitive Atoms, never sent to an AI provider, and never used to train anything. If you save the note you dictated, the saved text is treated exactly like any other capture described above.
For the website, the waitlist, and hosted accounts, we use:
In local-only mode, none of these receives the cognition you capture in the product. For hosted accounts, Railway and Supabase store your captured cognition as described in §4.5 — scoped to your account, and used for no other purpose.
There is none in the product. No usage statistics, no crash reporting, no feature tracking, no session recording. Our continuous integration rejects code that introduces a telemetry SDK, so this is a structural property rather than a setting you have to trust us to leave alone.
Waitlist email addresses are kept until the beta programme ends or until you ask us to delete yours, whichever comes first. Relay routing labels are kept only for the duration of a connection (§3). In local-only mode, your captured cognition is kept by you, on your device, for as long as you choose.
In a hosted account, your captured cognition is kept in your account until the account is deleted at your request (§8). Connection access tokens expire after one hour; refresh tokens after 90 days, or immediately on revocation.
Depending on where you live you may have rights to access, correct, delete, export, or object to our processing of your personal data. For the waitlist, write to privacy@cognitiveatoms.com and we will action it.
For cognition stored in local-only mode, these rights are yours to exercise directly: you can read, export, or delete the store on your own machine without asking us, because we are not holding it.
For a hosted account, deletion is self-serve: delete your whole account from inside the product — Profile in the apps, or your account’s profile page on the web. §4.5 lists exactly what that removes, and the three things it does not. Prefer to ask us instead? Email privacy@cognitiveatoms.com from your account’s email address and we will complete it and confirm to you within 30 days.
Our sub-processors may process data outside your country. Where required, transfers rely on appropriate safeguards such as the European Commission’s standard contractual clauses.
Cognitive Atoms is not directed at children and we do not knowingly collect personal data from anyone under 13. If you believe a child has given us data, write to us and we will delete it.
Traffic between an AI app and your machine is end-to-end encrypted, with the key held on your machine (§3). Credentials stored on your device are encrypted at rest. Website traffic uses HTTPS.
Hosted-account traffic uses HTTPS, connection tokens are stored hashed rather than in the clear, provider credentials you connect are sealed encrypted at rest, and every record is scoped to your account at the database layer (§4.5).
No system is perfectly secure. If you find a vulnerability, please tell us at security@cognitiveatoms.com rather than disclosing it publicly, and we will work with you.
If we change this policy we will update the date at the top. If a change materially affects how your data is handled, we will say so directly rather than relying on you to notice.
Cognitive Atoms, Inc.
privacy@cognitiveatoms.com