Privacy Policy
SuperLoom gives you a personal knowledge store ("brain") that you connect to Claude, ChatGPT, or another AI client over MCP. This policy explains what we store, where, who else touches it, and how to get it back or delete it.
Two things up front: SuperLoom isn't built for health information. It asks every connected AI not to save any, and you can delete anything that does get saved (see Health data). SuperLoom also never reads your AI's own memory, chat history, or conversation summaries. It only stores what you or your AI explicitly choose to save to SuperLoom.
Privacy FAQ
Who can see my memory?
You, and the AI apps you connect. Each account's memory is a separate database file. Your sign-in token selects your file, and no request can name anyone else's. Our server relays your memory only to the AI app you are using; what that app reads is then handled under your AI provider's own terms. SuperLoom's administrators can also see the mail log described under What we store, and can issue a new sign-in link for your account, for example when you are locked out; each one is recorded in an audit log.
Is my memory encrypted?
Yes. Your account's memory is encrypted with its own key, separate from every other account's. That key is locked under a second, server-wide key, kept outside the database. Deleting your account destroys the live copy of your key right away. This step is called crypto-shredding. A backup made before you deleted your account can still hold a copy of that key for up to 37 days; see Retention and deletion. Once every such backup has aged out, no one, including us, can decrypt what the key protected. See Where it lives.
Does an AI read my memory on your server?
No. Your own AI does the reading and reasoning, on your own plan. Our server runs only a small local embedding model for search and sends nothing to an AI provider. See AI models.
Does SuperLoom read my AI's memory or chat history?
No. SuperLoom has no access to Claude's, ChatGPT's, or any other AI's own memory, chat history, or conversation summaries. The only way something enters your SuperLoom brain is if you or your AI explicitly save it through a SuperLoom tool call, such as superloom_save.
What can SuperLoom do in my Google account?
Only what you approve on Google's consent screen, one Google account at a time:
- Calendar: see events and free/busy times, and create or update events. It can't delete events.
- Mail: search and read Gmail, draft, send and reply to email. It can't permanently delete mail.
- Drive: read only the files you pick yourself through Google's file picker. It can't upload, edit, or delete files.
Google shows an "unverified app" warning for the mail permission, because Google has not verified that permission for SuperLoom. What we keep from a connected account is under Connected Google accounts, and how we honor Google's data-use rules is under Google API Services User Data Policy.
Can I take everything with me, or delete it?
Yes, both from your Account page. Export downloads all of it as markdown and JSON (see Exporting your data). Delete removes your account and memory from our live systems in one step; backup copies age out as Retention and deletion describes.
Is my data used for training or ads?
No. We don't train any model on your data, and the site carries no ads, analytics or tracking. See No ads, no tracking, no third-party requests.
What we store
- Your brain's contents. The pages, facts, links, todos, timeline entries, skills, and tags you or your connected AI save, and the records from any file you import, in your own per-account brain file, encrypted with your account's own key (see Where it lives). Deleting a page, or asking your AI to forget something, removes the page for good, along with its search data, tags, links, timeline entries, imported records, and the facts saved about it. A copy remains only in backups until they age out (see Retention and deletion).
- Search embeddings. For each piece of text in your brain, a list of numbers that lets search find related text (see AI models below). They are stored in your brain file and deleted with it.
- Account information. Your email address, account creation date, and last sign-in time.
- Sessions and credentials. Web session identifiers. For each AI client you connect, a SHA-256 hash of the access and refresh tokens we issued it, so the tokens themselves are never stored. If you sign in with Google, we keep only the SuperLoom session it creates; we store no Google token for sign-in.
- Device and session metadata. A short browser/OS summary and a coarse IP prefix (the first three octets of an IPv4 address, or the first 48 bits of an IPv6 address) per signed-in device, for the device list on your account page. Neither is used for any security decision.
- A mail log. Every email we send you or receive from you (sign-in links, support mail, trigger notices, and the weekly email) is logged with its delivery status so an admin can see what was sent and received. A sign-in link's content is never stored, only a placeholder notice that it was withheld. The log is kept until you delete your account.
- Anonymous suggestions. If you choose to send SuperLoom a product suggestion from inside your AI chat, it is stored without any link to your account: no account id, email, session, or device information is kept alongside it.
- Tool usage. For each SuperLoom tool call, we store the tool's name, the action or skill name it ran, when, how long it took, and whether it succeeded. For each chat summary, we store its category and outcome. We never store the queries, page text, or other content behind a tool call. A chat summary's descriptive labels, such as a repeated task you named, are saved to your own brain. Tool calls older than 90 days are rolled up to daily counts and the detail is deleted.
- Error records. When a request fails, we record which tool failed, the error category, and a short error message, for 30 days.
- Connected accounts. See Connected Google accounts below.
Connected Google accounts
If you connect a Google account so your AI can use your mail, calendar, or Drive, here is what we keep.
- A refresh token for each part you grant. Google issues one for Calendar, one for Drive, and one for Gmail: three separate grants, each revocable on its own, even though Calendar and Drive share the same SuperLoom app registration. We store each one encrypted with Fernet (AES-128 in CBC mode, with an HMAC-SHA256 check that detects tampering) under a key kept in a file on our server, outside the database.
- Short-lived access tokens, in memory only. When your AI asks for your mail, calendar, or files, the server uses the refresh token to get an access token from Google. It holds that access token in memory until it expires (about an hour) and never writes it to disk.
- The account's details. Its address, the label and tags you gave it, which parts are connected, and the permissions Google granted.
Disconnecting an account, or deleting your SuperLoom account, sends each stored refresh token to Google's revoke endpoint and then deletes our copy. We delete our copy even if Google does not answer. You can check or remove SuperLoom's access yourself at any time at myaccount.google.com/permissions.
Your AI reads mail, calendar events, and files live from Google through SuperLoom. The server passes the results to your AI and does not store them. Three things can keep part of what was read:
- The alert inbox. Every 5 minutes the server checks your connected accounts for changes and records a short summary of each new item: for mail, the sender's name and address and the subject; for calendar, the event's title and start time; for Drive, the file's name, and only for a file you already picked through Google's file picker (it never learns about a Drive change outside that set); for each, a link and Google's id for the item. Each summary is deleted automatically 7 days after it was recorded, whether or not you have read it. That delete runs every 5 minutes. We never store a message body, an event description, or a file's contents in the alert inbox.
- Triggers you set up. A trigger can save a matching summary into your brain as a todo or a note, where it stays until you delete it. It can email you a notice, which says only how many items matched. Or it can send the summary to a webhook address you choose, and what that receiver does with it is outside our control.
- What you ask your AI to save. If you ask your AI to save something it read, or approve its suggestion to, that becomes part of your brain.
Google API Services User Data Policy
SuperLoom's use and transfer of information received from Google APIs to any other app will adhere to the Google API Services User Data Policy, including the Limited Use requirements.
What we read, and when. With your permission, SuperLoom can read your Google Calendar events and free/busy times and create or update events; read, draft, send, and reply to Gmail; and read the Drive files you pick yourself through Google's file picker (SuperLoom does not browse or read the rest of your Drive). Most of these reads happen on demand, the moment your AI calls a SuperLoom tool on your behalf. Two parts of SuperLoom also read on a short cycle without a tool call in progress:
- Coaching. When you start a chat, SuperLoom briefly reads your connected calendar (events in the surrounding two weeks: title, time, attendees) on our server to build that day's coaching suggestion. Those details are held in memory for up to 10 minutes, to avoid re-reading Google on every message in the same chat, then discarded; they are not written to disk or any database table.
- Triggers. Every 5 minutes, SuperLoom checks your connected accounts for new mail, calendar changes, or changes to the Drive files you picked (never other Drive files; the Drive file picker is the only way a file becomes visible to this check) and records a short summary in your alert inbox (see Connected Google accounts above for exactly what is kept and for how long).
Google data read this way is never used to train any AI model (generative or otherwise), never used for advertising or ad targeting, and never sold or given to a data broker. It leaves SuperLoom's server only to relay it to the AI client you are using, or into your own SuperLoom brain if you or your AI explicitly choose to save it.
Where it lives
SuperLoom runs on a single virtual machine hosted by Google Cloud Platform, in the us-central1 region.
Encryption. Each account's brain is encrypted with its own key (AES-256-GCM). That key is wrapped under a separate master key, kept outside the database on the server's own disk. Deleting your account deletes the live copy of your account's wrapped key in the same step as the rest of the deletion. This step is called crypto-shredding. A backup made before the deletion, such as the daily disk snapshot or the on-machine database backup and its cloud copy (see Retention and deletion for how long each one lasts, up to 37 days), can still hold a copy of the wrapped key and the master key that unwraps it. Once every such backup has aged out, your brain's encrypted contents can never be decrypted again, by us or anyone else.
Backups:
- The machine's disk is snapshotted daily, and each snapshot is kept for 7 days.
- A daily backup of the database and every account's brain file is kept on the machine for 7 days, and a copy is uploaded to Google Cloud Storage and kept there for 30 days.
- Each brain file is also copied continuously to a Google Cloud Storage bucket, so a brain can be restored to a recent point in time.
Backups exist to recover from data loss. They are not available as a copy you can request; use export for that.
Who else touches your data (sub-processors)
- Google Cloud Platform hosts the server, its disks, and the backups in Cloud Storage.
- Resend is the email-delivery provider. It sends sign-in links and account mail, and receives mail sent to superloom.io addresses. Account mail includes the weekly email, which is built from your brain: your first name, counts such as your open todos, and a suggestion for what to try next. Resend keeps its own delivery logs under its own policy.
- Google processes the OAuth sign-in exchange if you sign in with Google instead of a magic link, and, if you connect a Google account, the OAuth exchange and API calls your AI makes to read or act on your mail, calendar, or Drive on your behalf. We receive your email address and Google account identifier from these exchanges; we do not receive your Google password.
We do not use any other third-party processor.
AI models
No generative AI or language model reads your brain on our server, and nothing in your brain is sent to an AI provider by our server. Reading and reasoning over your brain happen inside the AI client (Claude, ChatGPT, or another MCP client) that you signed in to separately, under your own account with that provider. We relay your brain's data between our server and the client you are using.
To make search work, our server runs one small embedding model, nomic-embed-text v1.5, using llama.cpp on the same machine. It turns each piece of text you save, and each search you run, into a list of numbers that search uses to find related text. It cannot write or answer anything. It only accepts requests from the machine itself, and unless the faster-processing upgrade below is on, the text and numbers never leave it. We do not use your data to train any model.
Large imports are searchable by keyword as soon as they upload; turning them into numbers for search finishes in the background. Faster processing for large imports is an optional upgrade, off unless your company (or you, on a personal account) turns it on. With it on, imports over about 2,000 passages are processed by our own processing service in the same Google Cloud region as our server, running the same model. The text is encrypted while it waits, held in that service's memory only while it is processed, and deleted afterwards. It is never sent to an AI provider.
Voice in the web app
The web app can listen when you turn voice on with the voice control or its keyboard shortcut. Everything that touches sound happens in your browser, on your device:
- The microphone turns on only after you start voice, and only while the control says Listening. Your browser shows its own recording indicator the whole time. Voice stops on its own when the conversation or the period you chose ends, after a minute of push-to-talk, after 30 minutes with no speech, after 4 hours at most, when you leave the page, and on a phone when the page is hidden. Stopping voice also cancels anything still waiting to be saved.
- Speech becomes text on your device, using your browser's on-device speech recognition or a speech model you download. We do not use a browser's cloud speech service. Audio is never sent to our server or anyone else, and never stored.
- Spoken replies are made on your device, by a voice model you download or by your system's built-in voices.
- Our server receives only short text: the final words of something you save (after it is shown to you, with a moment to edit or cancel) or of a question you ask, which runs the same search as the Knowledge page. We do not log that text.
- No language model is involved. A fixed set of phrases decides whether you are saving a note, adding a todo, or asking a question.
Speech and voice models are downloaded only when you press a Download button. Your browser fetches the model files from Hugging Face (huggingface.co, which hands large files to its download servers under hf.co, such as us.aws.cdn.hf.co and cas-bridge.xethub.hf.co) and the speech runtime from jsDelivr (cdn.jsdelivr.net), at fixed versions, checks every file against a fingerprint we publish, and keeps them in your browser's storage. Those sites see an ordinary download request from your browser, including your IP address, and nothing else. Voice never saves something that looks like a password, PIN, key, or card number. Your voice choices (which recognizer and voice, and how long to listen) are kept in your brain with your other preferences.
Voice uses open-source software we serve unchanged from our own site while voice is turned on: ONNX Runtime Web (MIT), Transformers.js (Apache-2.0), kokoro-js (Apache-2.0), Silero VAD and the vad-web audio worklet (MIT and ISC), and phonemizer.js (Apache-2.0), which contains eSpeak NG. eSpeak NG is free software under the GNU General Public License, version 3 or later; it runs as its own file, only inside the speech worker. Its license text is at /js/vendor/phonemizer@1.2.1/COPYING and its source code, with how to build it, is at /js/vendor/phonemizer@1.2.1/SOURCE.txt.
No sale or sharing
We do not sell your data, and we do not share it with any party other than the sub-processors listed above, which exist solely to host the service and deliver email. We do not share your data for advertising purposes.
No ads, no tracking, no third-party requests
This site carries no advertising, no analytics script, and no tracking pixel. Every page on superloom.io loads its scripts, styles, images, and fonts from superloom.io itself; the fonts (Public Sans and Source Serif 4) are served from our own server. Viewing a page sends no request to any other server. Your browser only reaches another site when you follow a link to it, choose to sign in with Google, or press Download for a voice model (see Retention and deletion
Your brain's contents and account data persist until you delete your account or ask us to. Deleting your account removes: Deletion from the live database and brain file, and destruction of the live copy of your brain's encryption key, is immediate once run. What can remain afterward is only in backups made before the deletion, and only until each one ages out: our daily disk snapshot (kept 7 days, and it backs up the whole machine, so it can hold both the database and the separate master key file), the daily on-machine database backup (kept 7 days), that backup's copy in cloud storage (kept 30 days, or up to 37 if an object is replaced before it ages out), and the continuous copy of your brain file in cloud storage, which is deleted along with your account and otherwise keeps only recent history, pruning older history automatically. Any of these can hold a usable copy of your account's key alongside the data it protects until that specific backup ages out; after that, no one, including us, can decrypt it. We do not keep your data, or any copy of your key, longer than 37 days in any form. The same backup windows apply to anything deleted while your account stays open, such as a page you delete or an alert inbox summary. One record survives on purpose: an internal log entry noting that your account was deleted, which never holds your email or any other identifying detail beyond an opaque account id. Resend, our email provider, keeps its own delivery logs under its own retention policy. Export data on your account page downloads a zip file with everything we hold for you: Search embeddings are left out because they are computed from your text. There is no size limit. You can run up to six exports an hour. A paused account cannot sign in; ask support for an export. We recommend exporting before you delete your account, since deletion is permanent. SuperLoom isn't built for health information. We tell every AI you connect, including the ones built into SuperLoom's own tools, never to save health details. That includes a condition, symptom, medication, diagnosis, or an appointment's medical reason. It should keep only the non-health part, if there is one. For example, "dentist appointment Tuesday" is saved as "appointment Tuesday". SuperLoom's server does not check what gets saved, so this isn't foolproof: if you tell your AI to save health information anyway, or it does, SuperLoom will store what it's given. SuperLoom is not intended for protected health information and is not a HIPAA-covered service. If your organization is a HIPAA-covered entity, do not connect SuperLoom to systems that hold protected health information. If health information ends up in your brain anyway, delete it as soon as you notice; see Retention and deletion. We will update the effective date above when this policy changes and note material changes on this page. November 2, 2026: added a Google API Services User Data Policy section (the required Limited Use disclosure, what we read and when, and the server-side coaching and trigger reads), a note that SuperLoom never reads your AI's own memory or chat history, and updated the Google FAQ and Connected Google accounts to describe Drive access as only the files you pick through Google's file picker; added Where it lives's encryption details (each account's brain has its own key; deletion destroys the live copy immediately, and a copy in a pre-deletion backup ages out within 37 days), and reworded Health data and the top-of-page note: SuperLoom isn't built for health information and asks every connected AI not to save it, but SuperLoom's server does not check saves, so you can delete anything that gets saved anyway; and corrected Connected Google accounts and the Limited Use triggers bullet to describe Drive as its own grant (not shared with Gmail) that only ever sees the files you picked. October 27, 2026: large imports are searchable by keyword right away, and the optional faster-processing upgrade for large imports is described under AI models. October 21, 2026: added Voice in the web app, which explains that sound never leaves your device and which sites your browser downloads voice models from. October 19, 2026: the policy now names the local embedding model used for search, states that we store Google refresh tokens and how they are encrypted and revoked, describes the 7-day alert inbox and what triggers can save, covers every file in the export (which no longer has a page limit), lists the continuous brain backup, and states that the site makes no third-party requests now that fonts are self-hosted. Deleting a page now removes it for good. Questions about this policy: support@superloom.io (see Support).
Exporting your data
Health data
Changes to this policy
Contact