Skip to content
Gnum

Privacy policy

Last updated 26 October 2026

This policy explains which personal data we process when you visit gnum.io and when you create an account, why we do it, and which rights you have.

Gnum is in early access. Today it offers accounts and your own agent in Telegram and by email, which saves the conversations it can see, including the mail you forward or copy it in on, makes them searchable for you, and answers you with Claude, by Anthropic. Before Gnum does more, we will update this policy, as described under What comes next.

Who is responsible

Email tessa@gnum.io. Write to the same address with any question about your data.

Back to contents

When you visit gnum.io

Gnum runs on Cloudflare. To deliver pages and protect the site from attacks, Cloudflare processes technical data of every request: your IP address, the address you requested, the time, and your browser's user agent. Our application keeps request logs with this data for up to 7 days to find errors and abuse.

Legal basis: Art. 6(1)(f) GDPR, our legitimate interest in running a secure, working website. We use no analytics or advertising trackers, and the fonts come from gnum.io itself.

Back to contents

Your account

Sign-up and sign-in are provided by Clerk, which processes on our behalf:

Clerk sends verification codes and account emails from clkmail.gnum.io. To stop automated sign-ups, the sign-up form runs Cloudflare Turnstile, which checks technical data of your browser.

On your first sign-in, Gnum creates a random customer number and stores it together with your Clerk user ID in a database located in the European Union. It links your account to the data your agent will keep.

Legal basis: Art. 6(1)(b) GDPR, providing the account you asked for, and Art. 6(1)(f) GDPR for keeping accounts secure and free of abuse.

Back to contents

Connecting Telegram

Connect Telegram in your console opens Gnum's bot in Telegram. When you press Start there, Telegram sends us your Telegram user ID, name and username. We store your user ID with your customer number in the database in the European Union, and your name and username with the rest of your account's data.

When you create your agent, Telegram creates the bot in your Telegram account and tells us its ID, name and username. We store them and the bot's access token, which lets Gnum act as your agent. The token is stored encrypted, and Gnum never shows it to anyone.

Gnum also keeps a copy of your agent's picture, the one it has in Telegram, so that your console can show it, and the pictures you upload or make for it, the last 8, so that you can go back to one. They are stored apart from your other data, at Cloudflare, and only you can see them, signed in to your console. They go with your account when you delete it. A picture you upload, or make and use, is also sent to Telegram, which shows it as your agent's profile picture. To make a picture, Gnum sends an image model only a short description it writes from the style, mood and colour you choose or, when you describe the picture yourself, exactly your words, never your agent's soul, its name or anything else from your chats. Most pictures are made by Grok Imagine, the image model of xAI. When you describe the picture yourself, Gnum's decision model (on Cloudflare Workers AI) first reads your description, and only that, to choose the picture maker and to refuse sexual pictures of real people or of anyone who looks under 18. A sexual or suggestive picture is made by Venice (Venice.ai), which does not filter it, and everything else by Grok Imagine. When you ask your agent for a picture in your own words in Telegram, your words go to Anthropic with that reply, as any message does, and the words you gave for the picture go to the picture maker exactly as you wrote them. A picture made and not used is deleted after a day.

When you ask your agent for any other image in chat, up to 100 a month, the prompt for it goes to an image model: usually Flux 3 Klein, run for us by Cloudflare (Workers AI), or Grok Imagine (xAI) when the picture is not square and Cloudflare is not suited, and Venice (Venice.ai) for a sexual or suggestive prompt. Your agent writes that prompt from your request, or sends your own words exactly when you give a prompt yourself, and never adds your agent's soul or anything else from your chats. Gnum's decision model first reads that prompt, and only that, to choose the maker and to refuse sexual images of real people or of anyone who looks under 18. The image goes to the chat you asked in and is kept with that chat's files, apart from every other account's data.

Gnum also keeps a copy of the profile photo that Telegram or Slack shows for people in your chats, so that your console can show it beside their name. For Telegram, your agent's bot fetches the photo that a person shows people outside their contacts, as Telegram's own privacy setting decides, for people who wrote in a chat your agent saves and for you. For Slack, Gnum fetches the picture the person uploaded in that workspace, never Slack's letter pictures. Gnum fetches a photo only for the people a page you open in your console shows, and checks it again after a week. It keeps one small copy per person, a few kilobytes, in your account's own storage at Cloudflare, where only you can see it, signed in to your console. It is never sent to Anthropic or any other model. Gnum deletes the copy when Telegram or Slack no longer shows a photo, when it has not been confirmed for 30 days, with Forget person, when the period you chose for messages removes the person, when you merge two people, and with your account. The legal basis is Art. 6(1)(f) GDPR, your legitimate interest in keeping a searchable record of your own conversations, as for other people's messages in your chats.

Messages that people send to your agent reach Gnum through Telegram. Unless you let your agent talk with other people, what others write to your agent in private is never saved: Gnum keeps such a message only until your agent has answered it.

Your account's data, apart from the database entries named above, is stored apart from every other account, at Cloudflare, in a data centre near the place where you first opened your console.

You use Telegram under its own terms and privacy policy. Telegram delivers the messages between you, other people and your agent as its own service, not on our behalf.

Legal basis: Art. 6(1)(b) GDPR, providing the agent you asked for.

Back to contents

Saved conversations

Your agent saves the conversations it can see, so that you can read and search them in your console:

For each message Gnum saves the text or caption, who sent it (their Telegram user ID, name and username), when, which message it replies to, where a forward came from, earlier versions after an edit, and the reactions people leave on it (who, which emoji and when). Photos, videos, voice messages and files are not saved, only their type and details such as a file name or size, and the words in a voice message, a video message, a PDF, a Word document or a text file, as described below. Telegram's notes about the group, such as people joining or leaving, are saved too.

So that you and your agent can search your saved chats by meaning as well as by words, in any language, Gnum turns the text of each saved message, together with the two messages before it in its chat and who wrote them, into a list of numbers that stands for its meaning. An embedding model that Cloudflare runs for us (Workers AI) computes the numbers. Your account's own storage at Cloudflare keeps them apart from every other account's, each with only which of your chats the message is in and when it was sent, never its text. A search turns what you are looking for into such numbers the same way, and Gnum then reads the matching messages from your own saved chats. The numbers go when their message goes: when you delete it or the period you chose under Settings does, when you delete what a person wrote, when you delete a chat's messages, and with your account. The numbers of the two messages after it, which were made with its text, are then made again without it, and an edited message gets new numbers. Gnum does the same with the notes your agent writes about your chats and with the facts it keeps, each turned into numbers with the name of its chat or of the person it is about, so that your agent finds them by meaning too; their numbers go when the note or the fact goes, by any of the ways above or when you forget or delete it, and with your account. Your opinions and running jokes are turned into numbers the same way, kept in your account's own storage at Gnum with the rest of your data, and go at once when you change or remove one, and with your account. The messages of the chats you bring in from a file (see Chats you import, below) are turned into numbers the same way, each message that says more than a few words with the two messages before it that day and who wrote them, so that you and your agent find them by meaning too (your agent in your private chat only); for a chat of years this takes some minutes, in the background and the newest messages first, and meanwhile your agent finds the rest by words and says how far it has come. The numbers of a message go within minutes after it does, when you delete the import or the imported chat, delete what Gnum keeps about a person, stop saving a chat and delete its messages too, or the period you chose deletes them, and with your account; a big import's take a few minutes more, and the two messages after it that day are turned into numbers again without it. Before your agent replies to a message, Gnum turns the last few messages of the chat into such numbers once, to find what of the above is closest to what was just said, and gives your agent only what is close, within that chat's limits: in a group or someone else's chat, only that chat's own messages and notes and the facts you shared. In your private chat it also gives your agent the few messages of the chats you imported that are closest to what was just said, and what it wrote down about the people you mention and where your chats with them are, as counts and dates. These numbers are not kept.

So that you and your agent can find what was said in a voice message or a video message, and what a PDF, a Word document or a text file says, Gnum sends the recording or the file from the chats it saves (your private chat with your agent, your groups, the one-to-one chats you choose in Secretary Mode, the private chats other people start with your agent, and the attachments of the mail Gnum keeps) to a speech-to-text model and a document converter that Cloudflare runs for us (Workers AI), and saves the words that come back with the message. Your agent reads those words as part of the message, and like its text they go to Anthropic with a reply's request whenever your agent reads that message. The words go when their message goes, by any of the ways above, and with your account. Gnum does this for the messages it saves from 13 October 2026; voice messages saved before then, and those in chats you bring in from an export, have no words.

When you ask your agent to set up a reminder or a routine, a decision model that Cloudflare runs for us (Workers AI) reads the latest lines of your chat with your agent (up to six, each cut), and answers only a few questions about them: once or how often, and whether it is a reminder or something your agent does itself. Your agent then asks you instead of setting up what you did not mean. The model writes nothing, and nothing of it is kept. It also reads each message you write to your agent in your private chat with it in Telegram, while pictures can be made, and answers only whether you ask for a sexual or suggestive picture for your agent. If you do, Gnum makes the picture without your agent: your words go to the picture maker exactly as you wrote them, as for any picture you describe, and your agent writes no reply to Anthropic's model for that message (it reads the message later only as part of the chat, as it does any other).

While the setting “Set up reminders for me” is on, your agent decides in its own reply, from your conversation, when something you say you have to do deserves a reminder you did not ask for (it matters to you and you might forget it), and then sets it up for one time and tells you in a line; no separate model reads each of your messages for this. It is shown the titles of the reminders you have and of those you deleted lately, with your own words when you said why you deleted one, so that it does not repeat one or set up one of the same kind, and they go to Anthropic with that reply's request like anything else it reads. While the setting is on, the decision model also reads what you yourself write in a group your agent sees whole or in one of your one-to-one chats that Gnum saves, of the last day, with the two messages before it, and answers only whether you say you will do something for someone later (or for yourself), whether it is already done, and by when; Gnum keeps those answers with your message, and they go when it goes. For a reminder your agent set up on its own Gnum keeps why, and when you delete it, its title for 6 weeks, with your own words from the message in which you told your agent to delete it, up to 300 characters, so that it does not set it up again or one of the same kind.

A minute or two after someone writes to you in a group your agent sees whole, replying to you or naming your @username, or in one of your one-to-one chats that Gnum saves, the same model reads that message and the two before it, by who wrote them and never by name, and answers only whether it asks you to do something, whether it is you who is asked, whether it is already settled, and by which day. Gnum keeps those answers with the message, and they go when the message goes. When it does ask you, and the model that reads it for a reminder says it is yours to remember, your agent sets up a reminder for it on its own while that setting is on, and tells you; otherwise it mentions it to you once. Without a key, at your monthly limit, or when that request fails, nothing is set up on its own: the request is held, and the Routines page's Noticed lately says why. A reminder set up from what someone asked you quotes their words, and goes when their message goes.

Before that, a decision model that Cloudflare runs for us (Workers AI) reads each message that someone else writes in one of your one-to-one chats that Gnum saves, and each message in a group your agent sees whole that replies to you or names your @username, and answers only whether it may name something you have to do or remember. It does the same for each message you write yourself in those chats, to see whether you say you will do something later. When it may, Gnum sends Anthropic (Anthropic, PBC) that message, the four messages before it and the two after it by who wrote them, the writer's name as you know it, today's date and your time zone, your language, and the titles of your reminders with how each went (those you asked for yourself, those it set up on its own that came and stayed, those you have, and those you deleted lately, the last with your own words from when you deleted them, cut to 300 characters, when you gave a reason), so that a model of Anthropic's can say whether the thing is yours to remember, whether it is the same kind of thing as one you turned down, give its title in your language, and say the day and the time it names. A reminder is set up only when that answer says it is yours and is sure, and it goes when the message goes, as above. Nothing of it is sent while “Set up reminders for me” is off, at your monthly limit, or without a key, and the answer is kept only as the reminder or as why no reminder was made.

Groups, one-to-one chats and other people's private chats with your agent include messages from people who never signed up for Gnum. We process them on behalf of you, the account holder, who decides which groups to add your agent to and whether it talks with other people; tell the group before you add it. People in those chats can ask you, or us at tessa@gnum.io, to delete what was saved about them.

Legal basis: Art. 6(1)(b) GDPR for your own messages, the service you asked for, and Art. 6(1)(f) GDPR for other people's messages in your chats: your legitimate interest in keeping a searchable record of your own conversations.

Back to contents

Chats you import

With Bring in earlier chats in your console you can add chats from before: you export one chat, or your whole account, yourself in Telegram Desktop, and your browser reads the file on your computer, a piece at a time. For a whole export your browser reads the Telegram user ID in your account information, when the export has it, to check that the export is yours (without it, a one-to-one chat is brought in only when you wrote in it), and nothing else of it: your name, phone number, username, contacts, sessions, web sessions, profile pictures, other data and the chats you left are never read. It lists your personal chats and private groups, and you tick the ones to bring in; nothing of the others is sent. For the chats you tick, your browser sends Gnum only the words of each message, who wrote it (their Telegram user ID and the name the export shows), when, which message it replies to, whether it was edited, where a forward came from, Telegram's notes about the chat (someone joined, a new title) and, for a photo, voice message or file, its kind and details such as a file name or size. Photos, videos, voice messages and files, phone numbers, shared contacts, locations and the paths of files stay on your computer and are never sent to us.

For Telegram, your browser sends the chats you ticked once, and you can close the page when it says it is done. Gnum checks each part of a chat as it arrives and keeps it aside, in your account's own storage with the rest of your data, until it is brought in, usually within minutes and, for a whole account, up to about half an hour, one part after another; the page shows how far it got. What waits is only the fields described above, as Gnum checked them, and it can include messages that will not be kept in the end (no room is left, or they were imported or saved already). It is deleted as it is brought in, when you delete the import or the imported chat, when you delete what Gnum keeps about a person (which also stops every import still waiting), and within a day if sending was cut off. While it waits it is not in your download of your data or in the copy of what Gnum keeps about a person. It is never sent anywhere else, and nothing in it starts a message to anyone.

You can also bring in a WhatsApp chat: you export it yourself in WhatsApp with Export chat, Without media, which gives a .txt file on an Android phone and a .zip file on an iPhone, and your browser reads it on your computer, a line at a time; of a zip it opens only the chat file and never a photo, a video or any other file in it. For the one chat you choose, your browser sends Gnum only the words of each message, who wrote it, when, whether it was edited and, for a photo, video, voice message or file, its kind and nothing else: not the file's name, a shared place or a link to one. WhatsApp's file shows each author as the name your phone has for them or, for someone who is not in your contacts, as a phone number, shows times without a time zone and does not say which author you are, so Gnum asks you which author is you, in which time zone the phone was and, when the dates read two ways, which comes first. A phone number is sent to Gnum only so that it can be turned into a mark that tells it is the same person in your other WhatsApp chats. Gnum never stores the number, and does not write it in any log: it keeps that mark, and shows and keeps only the number's country code and last four digits. A name is a person of that one chat only and is never matched to anyone, whatever the name. These people are listed on your People page under From WhatsApp, where you can download what Gnum keeps about each and delete it; they are not contacts, unless you say that one of them is someone from your Telegram chats. Only you can do that, by choosing the person on People: Gnum never links anyone by a name or a number. A linked person's WhatsApp messages are then treated as that Telegram person's: they show under their name, count toward how well your agent knows them, follow your mark on them as a contact or a stranger, and are deleted with them if you delete what Gnum keeps about that person. Unlinking puts them back. Under Bring in earlier chats you can also change which author of a WhatsApp chat is you. The same chat again later adds only what is new.

Imported chats are kept apart from the chats your agent saves. You read and search them in your console, where a chat your agent saves too shows what you imported of it as its earlier messages, and the days people wrote in them count on your People page. Gnum also keeps the name a Telegram export shows for each person with their Telegram user ID, which is the name you saved them under when they are in your contacts, and shows it next to the name Telegram shows on your People page and with their imported chats; it never uses a name to tell who someone is. They are in your download of your data and in the copy of what Gnum keeps about a person. Your agent never replies in them, and nothing in them starts a message to anyone. In your private chat with it, your agent can look into them while it answers you, as it does in your saved chats; when it does, by words and by meaning, what it finds goes to Anthropic with that reply's request, and to nobody else, and so do the few imported messages Gnum finds itself, before the reply, to be closest to what you just said. To find a message by meaning, Gnum turns the text of each imported message that says more than a few words, with the two messages before it that day and who wrote them, into a list of numbers that stands for its meaning, as it does for a saved message (see Saved conversations, above): the embedding model Cloudflare runs for us (Workers AI) computes the numbers, in the background and the newest messages first, and your account's own storage at Cloudflare keeps them apart from every other account's, with which of your imported chats the message is in and when it was sent, never its text. The people in an imported chat never signed up, and the numbers made from their messages are deleted with the messages, as below. For the chats you brought in before this existed, the numbers are made in the background in the minutes after, an hour or two for the very largest, the newest messages first, a little at a time so as not to crowd out anyone else's. It never looks into them in a group, in someone else's chat with it, in a one-to-one chat it answers for you or in a message it starts on its own. Gnum keeps up to 2,000,000 imported messages for you, enough for a whole Telegram account. A whole account can mean years of messages from hundreds or thousands of people who never signed up: Tick all brings in every chat in the file, so bring in only the chats you want your agent to have, and delete any imported chat later. The amount does not change your agent's lookups for each reply, nor what Sleeps on it reads of your own words in the other chats you imported (up to 1,200 messages in all, as below); your one-to-one chats, though, are read whole, both sides, a little each night (see Your agent's replies, below), so more of them means more requests over more nights. A big import takes a few minutes to delete, about 7 for 2,000,000 messages. What Gnum keeps about a person you delete disappears from view at once, and their imported messages are gone within minutes. When you choose a shorter period for saved conversations, the older messages of a big import are removed within about half an hour, and then every day.

While Sleeps on it is on, your agent also learns a little from your own messages in the chats you imported. Once a night, and when you ask your agent to think things over, Gnum sends Anthropic up to three small requests, each with up to 40 of your own messages from the chats you imported other than your one-to-one chats, which are read whole, both sides (see Your agent's replies), the newest first (each cut to 300 characters, from the last 12 months of what you imported and up to 1,200 messages in all). They are only the words you wrote yourself, as plain text: never what other people wrote, never a forwarded message, a note of Telegram's about the chat, a photo or file, or a command. The request carries the facts your agent already has, the patterns it is watching for and what you rejected, and it answers with the same kinds of facts and patterns as the ones your agent notices in your private chat (see below), which Gnum checks against your words before it keeps them; it keeps no promise, question or plan from old messages. A fact or a pattern noticed in an imported chat is shown on your Digital twin page and under Memory with the chat and the day it comes from, and goes when the messages it rests on go: when you delete the import or the imported chat, delete what Gnum keeps about the person you talked with in a one-to-one chat, stop saving the chat with the same Telegram chat ID, or the period you chose for saved conversations passes. It never replaces a fact you told your agent. Learning is not paused by your monthly limit: it is metered and shown in Activity, and it goes on.

An imported chat includes messages from people who never signed up for Gnum and were not told. We process them on behalf of you, the account holder, who decides which chats to bring in; tell the people in a chat before you bring it in. People in an imported chat can ask you, or us at tessa@gnum.io, to delete what was imported about them.

Legal basis: Art. 6(1)(b) GDPR for your own messages, the service you asked for, and Art. 6(1)(f) GDPR for other people's messages: your legitimate interest in keeping a searchable record of your own conversations. We keep it small: only the words and who wrote them, nothing that is not text, the period you chose for saved conversations applies, and everyone can be forgotten.

Back to contents

Your email

Your agent has its own email address: a random one, and one of your own choosing if you want it, as long as nobody else has it. Mail to either address reaches your agent. Cloudflare's email routing receives it for us and refuses spam and mail that fails its sender's checks. Gnum keeps:

Gnum does not keep mail it drops: a bounce, an out-of-office or other automatic reply to mail your agent sent, mail that names none of your addresses, mail over the daily limit, mail too large to read, and mail that came while you had stopped taking mail in. For that mail Gnum keeps only counts, which the Email page shows for the last seven days. Confirmations from your mail provider, such as Gmail's code for forwarding, are passed to you: your agent tells you about them in your private chat, and the Email page shows them. The mail itself is kept with its sender, and what Gnum keeps of a confirmation is deleted after seven days.

Your agent sends mail only when you approve it: each mail it proposes is shown to you first, with its exact text, and nothing goes out until you press Send it. There are three exceptions: its answers to mail you write to it, sent to you; its answers in a thread it is addressed in, to the people in that thread; and, while you let it talk with anyone who writes to it in private, its answers to those people, within daily limits. It never writes as you.

Mail is kept as long as your saved conversations: as long as your account exists, or for the period you choose under Settings. Attachments go with their mail, and a stored file goes once no kept mail holds it. You can delete mail as you delete a chat or a person: delete what Gnum keeps about a person on People, which takes their addresses with them; stop saving a chat under Your chats and tick Also delete; or stop taking mail in on the Email page, which keeps no further mail and drops your forwarding address, while what is kept stays until you delete it. Switch off your forwarding in your mail settings too. Deleting your account under Your data deletes all of it at once.

Mail reaches and leaves Gnum through Cloudflare's email routing and email sending, which process it on our behalf. Your mail provider forwards your mail under its own terms. What your agent reads of your mail, to answer you or to learn, goes to Anthropic (Anthropic, PBC) with that reply, as described under Your agent's replies.

People who write to your agent's address are not your contacts and did not sign up for Gnum. Tell the people in your mail before you copy your agent in. They can ask you, or us at tessa@gnum.io, to delete what was saved about them.

Legal basis: Art. 6(1)(b) GDPR for the mail you forward or write to your agent and the answers you asked for, and Art. 6(1)(f) GDPR for other people's mail in your chats, as for saved conversations.

Back to contents

Your agent's replies

Your agent answers you in your private chat with it, and in a group when someone mentions it or replies to it and, where it sees every message, when someone calls it by name or carries on a conversation with it. If you let it join group conversations on its own, it also speaks up unprompted in groups where it sees every message. If you let it talk with other people, it also answers in the private chats they start with it, and tells you in your private chat when someone starts a conversation and, about an hour after it goes quiet, what they wanted and what waits for you. It answers in your one-to-one chats under Secretary Mode only if you switch that on in Settings, and then only people who wrote to you and got no answer from you for 10 minutes; it tells them it is an AI assistant. When it drafts a message in one of those chats for you to approve, it writes the draft as you would say it, from what it knows of you and how you write to them; the privacy check below reads that text as it reads a reply, and you see and approve the exact text before anything is sent.

To write a reply, Gnum sends Anthropic (Anthropic, PBC) the persona you set, up to five of the example lines you chose for your agent (below), the chat's latest messages (up to 30) with their reactions, and whatever your agent looks up in your saved chats while answering, and, in your private chat with it only, in the chats you imported (see Chats you import): in your private chat with it, any of your chats; in a group, only that group; in someone else's private chat with it or in a one-to-one chat it answers for you, only that chat. The reply comes back to Gnum and goes out through Telegram, and Gnum saves it in that chat. While a conversation with your agent in a chat goes on, Gnum also sends Anthropic the conversation's earlier turns again, as it kept them, so that the reply carries it on; a chat that is quiet for a few hours starts a new conversation, with the persona as you have set it by then.

You can tell your agent what to keep private: names, topics or facts that it never shares with anyone but you. Before a reply to anyone else, in a group or in a private chat, goes out, Gnum sends Anthropic a second, small request with your list, the chat's latest messages and the reply, to check that the reply does not tell any of them, in their words or in others. A reply that would is written again once, and held back if it still would; so is one that could not be checked. You are told in your private chat when a reply was held back. The list and the reply go to Anthropic only for this check, and only while you keep something private; each check counts toward your monthly limit.

Before your agent replies in a chat, or decides whether to join a group conversation, Gnum sends Anthropic a small request with the new messages it reads there, from anyone: your own words, what you forwarded and what others wrote since it last spoke (up to 20 messages and 8,000 characters). It answers only three yes-or-no questions, in whatever language the messages are in: whether they tell of hard news, such as a death, a serious illness or a lost job, of a crisis, such as not wanting to live, or of a health matter. When they tell of hard news or a crisis, your agent leaves out jokes and anything else it could bring up for the rest of that conversation. Before your agent shares something of its own with you, a request of the same kind reads what each thing it could share rests on: your messages to it, its notes about your groups, your plans, an entry of what it read in its own time (without its links) and, for a summary of a group, the group's latest messages; it shares nothing that tells of any of the three. The answers are not kept, except the fixed word for a reply described under What we keep, and each request counts toward your monthly limit.

You can mark a reply your agent sent as a good line and choose it as an example of how your agent sounds. Gnum then keeps a copy of that line, with whom it suits, its language and its mood. A conversation with your agent takes a few of the lines you chose that suit the people in it, and sends them to Anthropic with the persona. A line can come from a chat with other people, so choose lines that give nothing private away.

You can give your agent opinions: a topic, what it thinks and how strongly. When what people just said in a chat brings up a topic, up to three of the opinions on it go to Anthropic with the persona and the chat's latest messages, so that your agent says what it thinks. Anyone your agent talks with can hear it, so write opinions you are happy to have shared.

You can give your agent running jokes and references: the words that bring one up and what it is. When what people just said in a chat brings one up, and your agent has not brought it back in the last two weeks, one of them goes to Anthropic with the persona and the chat's latest messages, so that your agent can bring it back once. A joke it brought back lately goes along too, so that it leaves it alone. Anyone your agent talks with can hear it, so write running jokes you are happy to have shared.

Your agent also keeps facts you tell it in your private chat with it, such as where you live, the name of someone close to you, or anything you ask it to remember: what the fact says, the message it came from and when you wrote it. Gnum reads them out of your own messages itself, and what you forward to your agent is never kept as a fact. What other people tell your agent about themselves, such as where they live, is kept apart as their own words, with their name, the message it came from and whether they are a contact, a stranger or another agent, and never as a fact about you. A fact stays between you and your agent unless you choose, one fact at a time, to let it use that fact with anyone it talks with; the same goes for what others said. When your agent looks a fact up while answering, the fact goes to Anthropic with the rest of the reply's request. You can see, share and delete each fact under Memory in your console; a fact is deleted with the message it came from, and with the person when you delete what someone wrote. When a fact is a birthday you told your agent, your own or one of someone close to you, your agent uses it, in your private chat only: to wish you well on your own birthday, once a year, and, while Let my agent message me first is on in Settings, which it is by default, to write you one short wish that day if you have not written yet; and to ask you once, a few days before the birthday of someone close to you, whether you want a reminder, which it keeps only if you say yes. Gnum reads your answer itself, and a birthday that your agent only noticed, or that someone else told it, is never used. Neither a wish nor a reminder is ever sent to anyone for you without your own words.

For each person who writes in a chat Gnum saves, Gnum works out on how many days they wrote there, counting each day once, from the saved messages themselves, so the number goes down when you delete them. For people you mark as contacts you can also add a tone and a few words for how your agent talks with them, and a list of what is still open between you. When your agent answers a contact, a short note with these goes with the rest of what is sent for the reply, described above. When you delete what Gnum keeps about a person, these go too.

Ask me questions so it gets to know me and my people in Settings is on by default, and while it is on your agent now and then asks you in your private chat about someone who keeps turning up in the chats it saves: who they are to you (family, a partner, a friend, from work, someone you know, someone you hardly know) and, for a contact, how it should talk with them. You answer by tapping. Gnum keeps what you tapped with that person: it is shown on the People page, goes into the copy you can download of what Gnum keeps about them and is never said to them or to anyone else. Once you have said who someone is, your agent may also ask how you know them, what matters to them these days and whether there is anything to remember about them; you answer by tapping or by replying to that question in your own words. A message you write after a question without replying to it is never taken as the answer on its own: the question gets one button, Save, and one press saves it. What you tell your agent about someone is kept as a private note about that person, with the day you told it, for 12 months, and goes sooner when you delete it on the People page, delete the message it was read from, or delete what Gnum keeps about the person. It is used only in your private chat with your agent, never in a reply anywhere else and never in a message to another chat that you did not see in full. When a message you write there names such a person, your agent's reply is written with who they are to you and what you told it about them (up to two people, five lines each, picked by Gnum itself from your own words) added to what goes to Anthropic with the rest of that reply's request; no request of its own is made for it, and nothing of it goes with a reply in any other chat. The person is not asked and does not see it, but it is part of what Gnum keeps about them, so it is in the copy of it that you can download. The question and your tap are saved in your private chat like any other message there, so your agent reads them when it answers you there, but what you tapped is never put into its replies in any other chat. A question names the person, and at most two of the groups you both write in, only in the message to you; the copy saved in your chat leaves the groups out. Family, partner, friend and work also mark the person a contact, which you can undo on the People page. If you press Ask me about them under a person on the People page, Gnum keeps the time you pressed with that person until the question is asked, you say not to ask about them or switch the questions off, or 7 days go by. For each question Gnum keeps which person it was about, the step, how it ended and when, with none of your words. When you delete what Gnum keeps about a person, their answers, those records and the saved copies of the questions about them go too. If that person writes in your saved chats again on 3 more days, your agent may come to know them again and ask you about them, unless you switch the questions off. Your agent words each of its questions to you, about your people, about yourself and to check what it worked out, in its own voice and language: when it asks one, Gnum sends Anthropic one small request with your agent's soul, how it talks with you, the language you write to it in lately and Gnum's own plain wording of the question, its buttons and what it says once you answer, never what you answered or the names, words, facts or days it asks about, which Gnum puts in itself. Gnum keeps those words with the question, for as long as it keeps the question. The request counts toward your monthly limit; without it, the question waits.

If you let your agent join group conversations on its own, Gnum sends Anthropic the same after each pause in such a group's conversation, so that your agent can decide whether to say anything, even when it then stays quiet.

Your agent also picks up things by living: a view it formed, a running joke, a change of mind. Once a day, when your agent spoke up or reacted in a chat that day, Gnum sends Anthropic the oldest saved messages of those chats that it has not read yet (up to 150 in one request, each cut to 300 characters, with who wrote them), what your agent has picked up so far and the changes you rejected lately, so that your agent can propose a few changes to what it has picked up. When more than 150 wait, a busy day, a second request in the same night reads on from where the first stopped (at most 2 requests a night), and what is still left is read on the nights after, for up to a week from the day it started, after which it is let go; nothing is skipped in between. When a view or a changed mind is left, Gnum sends Anthropic a second, small request with those proposals and the messages they cite, to hold back the ones that only echo your own views. You can ask for this from the Soul page too, which sends the latest 150 saved messages of the last 7 days instead. What happens with a proposal depends on Let it grow on its own under Settings, which is on unless you switch it off. While it is on, your agent takes on the small ones that pass the check itself, at most 2 in a week, and only ones that rest on what you and it said to each other, never on what only strangers said; after a change it tells you in your private chat what it changed and why, and you can undo each change for a week, after which it stays until you take it back on the Soul page. While it is off, each proposal reaches you in Telegram with the messages it rests on, and nothing changes until you approve it. With Let it grow on its own on, your agent may also move how it comes across (how warm, playful, brief or full of emoji it is) one small step, at most once in two weeks for each, never more than one step from where you set it; Gnum works this out from the corrections you make ("too long", "say more"), whether you answer its longer replies at length, your emoji, thanks and laughs in your private chat with your agent and your reactions to its messages (never from how long your own messages are), in its own code over messages it has saved, with no request to Anthropic, and tells you in Telegram like any other change, with Undo for a week. In the same way, with no request to Anthropic, your agent may keep a running joke that you laughed at or loved twice as one of its signature touches, and let go of a touch that you have not answered in a month, working it out from your reactions and replies to its messages in your private chat. Changes that were held back, and why, are listed on the Soul page and never sent to you in Telegram. Those chats can be your private chat, groups and other people's private chats with your agent, so a change can echo something said there: undo, or do not approve, what gives something private away, or switch growing on its own off to decide each change first. Gnum keeps the proposals, the changes your agent made on its own, the ones it held back and how many tokens it took and what it cost.

While Sleeps on it is on, your agent spends a little time of its own, at most once in 20 hours and outside the hours you are usually around, reading about one of its interests. Then Gnum sends Anthropic one request with your agent's soul, with every name of a person and every private item taken out of it, that interest and the first words of its earlier entries on it, and nothing from your chats, messages or facts. Anthropic searches the web and reads up to two public pages for it, so what the request holds can end up in a search, and the entry your agent writes, with up to three links, is kept in your account under Its own days on the Soul page until you forget it. It is in your export and goes with your account. With Also share things of its own now and then on, it may write to you first with such an entry from the last week, once, as something it read: the entry, without its links, goes to Anthropic in the small request that words the message, as any share does, and in the day's plan when the plan weighs it. The request counts as learning and not toward your monthly limit. With Let it grow on its own on, an interest that your agent wrote at least three entries about, the first over four weeks ago, and that you never took up in the last four weeks (none of your messages names it and none of its replies about it was answered) fades, and Gnum sends Anthropic one small request, which counts as learning too, with your agent's soul, the views it formed and the interests it has and let go of, to pick a new one in its place. It tells you in your private chat, with Undo for a week, and keeps the new interest only if it names no person your agent knows and none of your private items.

Your agent also keeps a short note about each chat it saves: your private chat with it, the groups it is in, whether it spoke there or only read along, your one-to-one chats under Secretary Mode, whether or not it answered in them, and the private chats other people start with it. Once a day, and when you ask your agent to think things over, Gnum sends Anthropic, for each such chat with at least six saved messages it has not read yet, one small request with the oldest of them (up to 60, each cut to 300 characters, with who wrote them) and, when more than that waits, up to two more requests for the same chat in the same night, three at most. When the note a request writes comes back longer than 400 characters, Gnum sends the same request once more with the note it wrote, asking for a shorter one, and cuts what is still too long at the end of a sentence. Up to ten chats are read in a night: your private chat with your agent first, then the chats your agent spoke in, then those a night left half read, then those with the most new messages; what is still left is read on the nights after, and a message older than 30 days is not read. So what the people in your groups and one-to-one chats write goes to Anthropic for these notes even where your agent never spoke. The request asks for a note that leaves out passwords, account and card numbers, anyone's health, money or family troubles, and anyone's religion or beliefs, political opinions, sex life or sexual orientation, ethnic origin, trade union membership and genetic or biometric data. Gnum sends nothing from your facts or any other chat. Your private chat with your agent is one of these chats: its note is written from its own messages alone, and only your private chat with your agent ever reads it. The note it writes, up to 400 characters with a few topic words and open questions, is kept with the list of the messages it was made from, and goes when any one of them goes: when you or retention delete it, when you forget the person who wrote it, when you delete or stop saving the chat, or when Telegram reports it deleted or edited. You can read and forget the notes under Memory in your console. Notes older than 30 days become one line for the week, built without a request, and the lines of weeks older than 26 weeks become one line for the month, built the same way, which is kept like a note: with the messages it was made from, until you forget it or one of them goes. Your agent also uses its notes when it replies. The request for a reply in a chat goes to Anthropic with that chat's own last two notes and any older note of that chat that what people just said brings up, and never with another chat's notes; the request for a reply in your private chat with your agent carries the latest notes of all your chats, the latest note about that chat itself among them. A note that names something you keep private is left out of replies to anyone but you. Gnum keeps how many tokens each request took and what it cost, 400 days.

Your agent also notices facts you plainly state about yourself. Once a day, and when you ask your agent to think things over, if you wrote at least three messages in your private chat with your agent since the last time, Gnum sends Anthropic one small request with the oldest of those messages that it has not read yet (up to 40, each cut to 300 characters; never a forwarded message, never anyone else's), with the facts your agent already has, the patterns it is watching for and what you rejected. When more than 40 of your messages wait, a busy day, up to three more small requests in the same night read on from there, four at most; what is still left is read on the nights after, and a message older than 30 days is not read. It answers with facts of a few fixed kinds (where you live or are moving, where you work, what you are called, a birthday, a diet or allergy, languages, the name of someone close, what you like or dislike), each with the words from your message that state it; with promises and questions you left open, each with your own words; with at most two things you say are coming up on a day, such as an interview on Thursday, each with the words that say it; with patterns in how you write, what matters to you, how you like things done and your routines, each as one sentence with the messages that show it; and with up to three things you say about who someone in your chats is or what you call them, each with the names your words use for that person and your words. Gnum keeps a fact only if those words are in your own message and nothing in them is doubtful, and refuses it otherwise; what you say about who someone is goes on their record as when you tell your agent in a reply (below), only if those words and names are in your own message, and never against a No you gave. A noticed fact is private, shown under Memory with your own words next to it, goes with the message it came from, goes after 90 days if no reply ever used it, and never replaces a fact you told your agent: when your words disagree with one, your agent asks you first. Forget removes a noticed fact, and Gnum then remembers for 180 days what you rejected, as the same words, so that it is not noticed again. A promise or question is kept only if those words are in your own message and say it plainly, and it goes on your list of things left open under Settings, with your own words, marked as found by your agent. Something coming up is kept only if those words are in your own message, give a day within 60 days in your saved time zone, and hold no doubt, no order, nothing about health or a loss and no hard news, in the words or the message around them; Gnum reads the day from your words itself, and without a saved time zone it keeps none. The day after, your agent may ask you once how it went, in your private chat; it fades 3 days after its day, and Close removes it. Your agent may bring it up in your private chat, you can close it there at any time, it goes with the message it came from, and one that nobody brought up fades after about two weeks. When your words disagree with a fact you told your agent, your agent asks you about it in your private chat, and your answer settles it. A pattern is shown on your Digital twin page only once at least three of your messages on two different days show it, and is used by your agent only after you keep it, there or when it asks you in your private chat, and then only in your private chat with it, never in a group or in anyone else's chat. Taking it back, or Not this, removes it. The request goes under the same terms as the others.

Your agent also learns about you from what you write to other people, from your own words only. Once a day, and when you ask your agent to think things over, if you wrote at least three messages in your other saved chats since the last time, the groups and your one-to-one chats under Secretary Mode, Gnum sends Anthropic one small request with the oldest of those messages that it has not read yet (up to 40, each cut to 300 characters, from those chats together and without saying which chat each is from), with the facts your agent already has, the patterns it is watching for and what you rejected. When more than 40 wait, up to three more small requests in the same night read on from there, four at most; what is still left is read on the nights after, and a message older than 30 days is not read. It sends only the words you wrote yourself, as plain text: never what anyone else wrote, never a forwarded message, a deleted message or your agent's own messages. Of what comes back, Gnum keeps only facts you plainly stated about yourself and patterns in how you write and do things, with the same checks as above, and never a promise, a question or something coming up, which stay between you and the people you said them to, nor your diet or an allergy; the request is asked to note nothing about anyone's health, religion or beliefs, politics, sex life, ethnic origin or trade union membership, and a pattern that names a health matter is not kept. What it keeps goes with the message it rests on, when you edit that message, when you stop saving the chat, and, for a one-to-one chat, when you delete what Gnum keeps about the person you wrote to, though your own messages stay in the chat. The request goes under the same terms as the others.

Your agent also gets to know the people you write with one to one, and learns about you from what they write to you, but keeps nothing of what they say until you say it is right. While Sleeps on it is on, once a night, and when you ask your agent to think things over, Gnum reads both sides of your one-to-one chats: those under Secretary Mode that it saves, and the one-to-one chats you brought in from Telegram, a saved chat and the imported chat it continues as one conversation. It sends Anthropic small requests of up to 150 messages of one chat each (each cut to 300 characters, with who of the two wrote it and the day), with your first name, the other person's name and the names you said they go by, what your agent already knows about the two of you and what you said is not right: a chat's new messages first, then its earlier ones, the newest first, up to 12 requests a night, and as many each time you ask, across all your one-to-one chats, about a cent each. It is not paused by your monthly limit, so a long chat is read over many nights (one of 190,000 messages over months), and your Digital twin page shows how far each chat has come. Only plain text messages of the two of you are sent: never a forwarded message, a photo or file, a deleted message, a command or your agent's own messages. Gnum checks what comes back against the message it cites and keeps only a few kinds of facts stated plainly in that message: where someone lives, is moving or works, what they do for work, what they like to be called, a birthday, the name of someone close to them, what they like or dislike and what they do regularly, about the other person in their own words or in yours to them, or about you in theirs; never a guess, a joke, a question or something that would or could be, nor anything about anyone's health, diet, religion or beliefs, politics, sex life or sexual orientation, partner, ethnic origin, trade union membership or genetic or biometric data, nor a death, an illness or another loss. What you say about yourself there is noticed as in your other chats (above). Everything else waits for your yes: it is never used in a reply, never shown to anyone and never sent anywhere until you say it is right, when your agent asks you (below) or on the person's entry on your People page. Yes keeps it: a fact about the other person is used only in your private chat with your agent, when you name them or it looks them up, and is kept for 12 months after your yes; a fact about you joins your other facts. No forgets it, and Gnum remembers for 180 days, under that person, that you said no, so that it is not put to you again. What waits is kept for up to 12 months. What waits and what you said yes to go with the message they rest on: when you or the period you chose delete it, when Telegram reports it deleted, when it is edited and no longer says it, when you delete the import or the imported chat, when you stop saving the chat, and when you delete what Gnum keeps about the person, which also takes your no's about them. The copy of what Gnum keeps about a person includes what you confirmed about them and what waits, with the words it rests on. Gnum keeps how far it has read each chat (no words) and how many tokens each request took and what it cost, 400 days.

Your agent also writes down a short picture of each person you talk with, so that it knows who they are when you ask about them: who they are to you, what you talk about, what they are up to lately and the language they write in. While Sleeps on it is on, once a night, and when you ask your agent to think things over, Gnum sends Anthropic one small request for each of up to 8 people, about a cent each, only for someone your chats show on at least 3 days: first the people you write with one to one, marked as a contact or said who they are to you, then the others, each those your chats show on the most days first. A request carries what that person wrote in the chats Gnum saves and in the chats you brought in (their newest 80 plain messages, each cut to 240 characters), what you wrote to them in your one-to-one chats with them (your newest 40), never what you or anyone else wrote in a group, the notes your agent keeps of their own private chats and those of your private chat with your agent that name them, never a group's note, the facts about them that hold (what you told your agent, what you confirmed from your chats, what they said about themselves) and your own facts that name them, who you said they are and the names you call them, and the lines you removed before. A person's picture is written again only once 40 new messages came from them or from you to them one to one, or you brought in your one-to-one chat with them again, and no sooner than a week after the last. It is not paused by your monthly limit, so the people already in your chats get theirs over a few weeks. Gnum checks every line before it keeps it: one plain sentence that rests on at least one message, note or fact the request carried, with no link, no order to your agent and nothing about anyone's health, religion or beliefs, politics, sex life or sexual orientation, ethnic origin, trade union membership, or genetic or biometric data, nor whom they are with unless it is you, nor hard news. A line about someone goes with the message, note or fact it rests on: when you or the period you chose delete it, when Telegram reports it deleted or it is edited, when you forget the note or the fact stops holding; every line of the picture goes when you stop saving a chat it read, delete the import it read, or when the period you chose deletes a message it read; the line of who they are to you goes when you say who they are; and the whole picture goes when you delete what Gnum keeps about the person. Your agent uses a picture through its lookups, and in your private chat with it also before it replies to you when your words name the person: in your private chat with it, any of them, and in another chat only one written from that chat's own messages and notes alone, through its lookups, without a line that names something you keep private; it is never part of its instructions. Each person's entry on your People page shows their picture, with what each line rests on, and you can remove any line, which Gnum remembers for 180 days so that it is not written again. The copy of what Gnum keeps about a person includes their picture, with what each line rests on.

When you ask your agent in your private chat to go through whole chats for you, such as everything someone wrote to you over years, it can start a job that reads them in the background and comes back with what it found. It starts one only on your own words in that chat, never because of something forwarded or written by someone else, and never in a group, in someone else's chat with it or in a message it starts on its own; one job reads at a time. You choose the chats, those Gnum saves and those you brought in from a file, and a period if you like; a chat you stopped saving and a group you left are never read. Before it reads anything, Gnum works out what reading them will cost, from how many messages they hold and what each request costs, and your agent tells you in the same reply. The job then sends Anthropic small requests, a few seconds apart, each with what you asked to look for and up to 200 messages of one chat, the newest first: both sides, other people's words included, each cut to 300 characters, with who wrote it and the day, never a forwarded message, a photo or file, a deleted message or your agent's own messages. Gnum keeps only what comes back that rests on a message the request carried, with words that are in that message letter for letter, and never anything about anyone's health, religion or beliefs, politics, sex life or sexual orientation, ethnic origin, trade union membership, or genetic or biometric data, nor hard news. What it found is kept only until your agent has told you, in your private chat, in a message of its own written by one more request to Anthropic with your persona, what you asked, how far it read and what it found, with who wrote it, where and when; when that request cannot be made, Gnum sends you what it found itself. That message is kept like the others in your chat. Nothing a job finds becomes a fact, a note or anything else your agent keeps, and until you are told, it goes with the message it rests on: when you or the period you chose delete it, when Telegram reports it deleted or it is edited and no longer says it, when you delete the import or the imported chat, when you stop saving the chat, and when you delete what Gnum keeps about the person who wrote it, or about the person whose one-to-one chat it is in. A job stops when it has read everything, at its cap of $2.00, or when you tell your agent to stop, and you can ask your agent how far it got; each request is metered and shown in Activity, without counting toward your monthly limit. While a job has not told you yet, the copy of what Gnum keeps about a person includes what it noted from their messages. Gnum keeps what you asked to look for, the chats and the period, how far the job read, how many things it noted and what it cost.

On the Check tab of the Digital twin in your console, a test predicts your replies. About once a week, while Sleeps on it is on, and when you press Run it now, Gnum takes up to 40 moments from the last weeks in which you replied in your one-to-one chats, those under Secretary Mode that it saves and the ones you brought in: what someone wrote to you in a burst of up to 6 messages and what you wrote back (up to 3 messages, 600 characters), never a group, a forwarded message, a deleted message or your agent's own messages. For each moment it sends Anthropic one request that asks for your reply, with what your agent wrote up about you, your people and the latest months from what it had learned before the weeks that are tested began (your profile, who matters to you and what you did lately, as it would carry to your private chat), the other person's name and who they are to you if you said so before that moment, the last 12 lines of that chat before it (up to 2,000 characters), the words they wrote, and what your agent's memory finds in your chats and notes from before that moment, never anything dated after it. A second request, with the words they wrote, the lines before them, your real reply, the guess and your real reply to another moment put among the guesses on purpose, marks each from 0 to 4 for what it says and for how it is said, to see whether the marks can be trusted. Nothing goes to Telegram or Slack and nothing your agent keeps is changed; each request is metered and shown in Activity, without counting toward your monthly limit, and a run costs about a dollar. Gnum keeps each run with its moments by reference to your messages, not copies, the guesses, the marks and their reasons, the picture it was tested with, what it cost and how it ended.

Your agent also writes who you are: one account of you in six sections (what you put first, how you decide, what you care about now, your opinions, the patterns in how you live, and how you have changed). Once a reading of your whole chats and its write-up are done, and then about once a month while Sleeps on it is on, Gnum sends Anthropic one request with your profile and the lines you confirmed about the people who matter to you, your years and the last six months, up to 250 of the things that hold about you, up to 120 of the most important moments across the years, and a sample of up to 150 of your own messages (at least 40 characters each, cut to 300, from your one-to-one chats across the years and the people you write to: only what you wrote yourself, never a forwarded message, a deleted message or your agent's own messages), with the names of the people they involve but not who they are to you, the sentences you confirmed and the ones you removed. It is written whole each time, never added to. Each sentence shows on your Digital twin page what it rests on, with Confirm and Remove; one that names someone else waits for your yes, and one you remove is not written again. A second request writes it again from what your agent had learned before the weeks the prediction test tests, so that the test can tell whether it helps; that copy is never shown to you and never used in a reply. Each request is metered and shown in Activity, without counting toward your monthly limit, and a version with its test copy costs about half a dollar. When a version comes into use, your agent also tells you about it once in your private chat, with one more request to Anthropic that carries those sentences and the ones new and gone since the version before; it is metered and counted toward your monthly limit like a reply, and shown in Activity. The request goes under the same terms as the others.

Once a version of who you are is in use, Gnum sends Anthropic a small request, on a small model, with the sentences of it about what you care about, what you value and the patterns in how you act, and with the things your agent could bring up with you (an open loop, someone waiting for you with their name and the words they wrote, a thought it has, a share of its own) and the reasons it is about to pass on to you from another chat, so that it can score what matters to you; the answer is one number for each thing, and nothing of who you are goes anywhere else.

When your agent tells you what it learned about you (after you ask it to think things over, after a reading of your chats ends, and once a day in the evening when something changed, which you can switch off in Settings or hold off for three days with Not now), Gnum sends Anthropic one request with your agent's soul and how it talks with you, and what changed in its picture of you since the period began: the sentences of who you are that are new or gone, your corrections, the claims about you that you told it or confirmed, the things about other people that wait for your yes, and the moments it noted. Hard news is left out unless you ask for it yourself. The answer is one short message with no score in it.

With each version of who you are, your agent also writes how you write to your partner, to your family, to your friends, to people you work with and to the people you write to most, as one short paragraph for each. For each one Gnum sends Anthropic one request with up to 120 of your own messages to those people from the last year, sampled across it (only what you wrote yourself, never a forwarded message, a deleted message or your agent's own messages), the name of the person, or for a relationship the word for it (partner, family and so on), and the numbers it measured itself, such as how long your messages are and how often you use emoji. Your agent keeps up to five of those messages as examples of how you write, and shows the paragraph and the examples on your Digital twin page, where you can edit or remove it. Each of these requests costs about two cents, and a version has up to about thirteen of them, with as many for its test copy, on top of the figure above. They are metered and shown in Activity, without counting toward your monthly limit. The prediction test also reads the paragraph and examples for the person a moment is with, in the same way it reads the rest of who you are. Forget person deletes what is kept for them.

Your agent also writes who each of the people you talk with is: one account of them in five sections (how it began and since, where things stand, what matters to them, you and them, and lately). It does so for people it has a page about and whom you wrote with one to one in the last 90 days, right after a run of the picture that wrote pages, and once a night while Sleeps on it is on, for those about whom something was recorded or decided since, up to 8 people a run. For each, Gnum sends Anthropic one request with the lines of their page and whether you confirmed each, up to 250 of the things that hold about them and the people close to them, the years of your chat with them and its last twelve months, up to 80 of their most important moments, what was noted with them or about them in the last three months, and the sentences you confirmed and the ones you removed, with the names of the people involved but not who they are to you or to anyone. It is told to write nothing of the special categories and nothing of a death, a serious illness or another hard news. It is written whole each time, never added to. Each sentence shows on the person's page, under What it knows, what it rests on, with Confirm and Remove. Your agent uses a sentence in a reply or a lookup only once you confirmed it, or when everything it rests on is already yours (something you told it or confirmed, a line of their page you confirmed, or a moment of your chats) and it names no one else; any other sentence waits for your yes, and one you remove is not written again. In your private chat with your agent, and nowhere else, these sentences take the place of the shorter profile and of the lines of their page. Each request is metered and shown in Activity, without counting toward your monthly limit. The request goes under the same terms as the others.

While Ask me questions so it gets to know me and my people is on, your agent also checks with you, one thing at a time and in your private chat only, what it worked out about you but you never told it: a pattern that shows on your Digital twin page, a fact of a personal kind it noticed in your own words (a birthday, a diet or allergy, the name of someone close, in your private chat with it, in your other saved chats or in a chat you imported), and a fact that may have changed; and what it picked up in your one-to-one chats about the other person or about you in their words (above), after your own things, with their words and who wrote them. It asks either in a reply to you, at most once a day and never while you seem stressed or after you shared hard news, or in a message of its own with Yes, No and Not sure, which Gnum writes itself, at most once a day and within the limits of Writing to you first. When it asks in a reply, the request for that reply carries the pattern's sentence, or the fact with the words from your own message it rests on and the day you wrote them, to Anthropic with the rest of that reply's request; no request of its own is made, and none is made for a message with buttons. Your next plain message in that chat settles a question in a reply: a short yes keeps it, a short no forgets it, anything else leaves it. Yes keeps a pattern as you would on the Digital twin page and makes a noticed fact one you confirmed; No removes it and Gnum remembers for 180 days what you rejected, as the same words, so that it is not noticed or suggested again; Not sure, or no answer, leaves it as it is to be asked again in 2 weeks, three times at most. A question with buttons and how you answered are saved in your private chat like any other message there.

While Ask me questions so it gets to know me and my people is on, your agent also asks you about yourself, one short question at a time and in your private chat only, about what it does not know yet: what you do for work, which city you live in, when its reminders should come, what you enjoy in your free time, what you do regularly and who is in your family. A topic is not asked while your agent has a fact about it, however it came to have it. You answer by tapping one of the answers offered or by replying to the question in your own words; a message you write after a question without replying to it is never taken as the answer on its own: the question gets one button, Save, and one press saves it. Your answer is kept as a fact you told your agent about yourself, with your words and the message they came from (none for a tap), like the facts above: private, used only in your private chat with your agent unless you let it use it with anyone, and kept until you delete it under Memory or delete the message. Not now leaves the topic for 2 weeks, Don't ask me this never asks it again, That was not an answer deletes the fact, and a plain no keeps nothing and does not ask again. It asks at most once a day, within the limits of Writing to you first, never within a day after you seemed stressed or shared hard news, and a question nobody answers only puts the next one off by a day. No request is made for it: Gnum writes the question and reads your answer itself. For each question Gnum keeps the topic, how it ended, when, and which of your messages or which fact answered it, with none of your words; the question and your answer are saved in your private chat like any other message there.

When you tell your agent in your private chat who someone in your chats is, or another name you use for them, in any language, such as “A F is Nastja, she is my wife”, it looks them up among the people of your chats and puts it on their record: who they are to you, as if you had answered its question about them (family, a partner, a friend and work also mark them a contact), and the new name as one more name of theirs. Gnum checks in code that your own words say it and hold the names, never a forward or anyone else's words. When a name you used is that person's and none is anyone else's, it is set at once; when none is theirs, or someone else goes by one too, nothing is set until you say so: your agent asks you in a message of its own with Yes, No and Not sure, which Gnum writes itself, at the pace of its other questions. Yes sets it and keeps the name you used as theirs; No keeps them apart, and that name is not put to you for them again. The title of a one-to-one chat you brought in, the name your Telegram or phone shows for that person, is one more name of theirs too. Your agent knows a person by these names in your private chat with it: when you name them, when it looks them up and when it searches your chats; People shows them with what each rests on and deletes them. Something you told your agent before about your wife, husband, partner or another relation is put on their record the same way, without a question when your words named that person and you write with them one to one; otherwise nothing is guessed. The person is not asked and does not see it. No request of its own is made for any of it.

Your agent keeps one list of the people in your chats, so that each is one person to it under every name and on every account, with who they are to you and to each other: your wife, her mother. While Sleeps on it is on, once a night, and when you ask your agent to think things over, Gnum reads who is who, and only when something new waits does it send Anthropic one request, about 5 cents: your first name, the people concerned (the names they show and go by, their accounts, who you said they are to you and to each other), the words you put first or last in your own messages to someone one to one, set off by a comma, that show in at least 3 of your newest 500 messages there (“Настя, ...”), each with up to 3 of those messages, and two people whose names are alike and who could be one person, such as Pablo in Slack and Pablo Inigo in Telegram, each with up to 3 of their own newest messages, every message cut to 300 characters, never a forwarded message, a photo or file or a deleted message. It answers which of those words is a name you call that person by, which two people are one, and who someone is to you or to someone else, each with the messages it rests on. Gnum checks that each message is one the request carried and holds the words quoted from it. A name you call someone by in at least 3 of your messages to them is kept on their record at once, and People shows it with your messages that show it; two people being one and who someone is to you or to someone else wait for your yes, which your agent asks in a message of its own with Yes, No and Not sure, at the pace of its other questions. Yes makes the two one person, as Same person does, or keeps who someone is; No keeps them apart, and it is not put to you again. When you said yes to something your agent picked up about someone close to the other person in a one-to-one chat (above), such as where their mother lives, that person is kept too, with no account of their own in your chats, as “their mother” or by the name it gave. Your agent's other readings of your chats (the notes, your own words, both sides of a one-to-one chat, the pictures of people and research) are told who is who in the chat they read, by Gnum's own numbers for those people, and say whom what they found is about by those numbers, reading a word like “mom” as the mother of whoever writes it. Each request is metered and shown in Activity, but learning does not count toward your monthly limit.

Your agent reads your chats from the first message by itself: about ten minutes after you bring chats in from Telegram, and then in its hourly check, your biggest one-to-one chats first (your partner's and the next five biggest, saved through Secretary Mode or brought in from Telegram), and then the rest of your chats, the groups you brought in from Telegram included. Your Digital twin page shows what it is estimated to cost and how far it got; Stop is always there, a stop holds until you choose Resume, and both readings are metered and shown in Activity without counting toward your monthly limit, so your replies keep all of it. Gnum sends Anthropic, mostly through its batch service (requests it answers within a day, at half the price), requests of about 20,000 to 40,000 tokens each: a stretch of one chat, both sides, other people's words included, each message cut to 600 characters with who wrote it and when, never a forwarded message, a photo or file, a deleted message or your agent's own messages, with the people of that chat (the names they show and go by, who they are to you and to each other). It answers with what happened and what holds about whom, each citing messages of that stretch with their exact words. Gnum keeps only what rests on messages the request carried, with words that are in them letter for letter, written by the person it says wrote them, and never anything about anyone's health, religion or beliefs, politics, sex life or sexual orientation, ethnic origin, trade union membership, or genetic or biometric data, nothing that reads as an instruction, and from a group nothing about anyone but you. A second, separate request checks each thing that would count, and each thing someone has one of at a time (where they live, where they work): it reads the messages around it (up to 12 on each side) with the people of the chat, and the thing without whom it was taken to be about, and Gnum counts it only when that check finds the same person, the same people with them and the same time, and that it is said as a fact. When two values of such a thing were said too close together for Gnum to tell which came first, one request with the two values and the words they rest on says how they fit. What is kept shows on your Digital twin page, each with the messages it rests on and how sure Gnum is of it, which comes from where it comes from (your own words, someone's words about themselves, or someone's about another), whether the check agreed and how often it was said, never from the model's own word; none of it reaches a reply yet, and what it holds about other people waits for you to go through it. When it is done, your agent tells you in your chat, in plain numbers, what it read and found.

Once a reading of your chats is done, and then once a night while Sleeps on it is on, your agent writes up what it found: each month of a chat, each year of it and each year of your life, your profile (who you are, as your chats show it) and a page about each of the people who matter most to you. Gnum sends Anthropic, mostly through its batch service, requests that carry only what the reading kept, never the messages themselves: the moments it noted (a sentence or two each, with the day and the people they name), the lines it wrote before about months and years, and the things that hold about you or about that person, each with how sure Gnum is of it, with the people of the chat (the names they show and go by, who they are to you and to each other) and the lines you confirmed or removed. It answers with lines in English that each cite what they rest on, and Gnum keeps a line only when what it cites was in the request and is about you or that person, and when it touches none of the special categories above and reads as no instruction. Right after a reading it counts against that reading's budget, which its estimate includes; at night against the night's own limits, which are not your monthly limit. Each line shows with what it rests on, on your Digital twin page or under the person on People; what it says about other people is used only once you confirm it there, line by line, and a line you remove is never written again. In your private chat with your agent, and nowhere else, it reaches the replies: when a conversation starts, the lines of your profile it is fairly sure of, the people who matter most to you with who they are to you (a line about one of them only once you confirmed it) and the latest months of your chats; before a reply, what you confirmed about a person your words name and a few moments with them, and what a time your words name held, which the small request that already reads every reply for hard news also reads, with today's date; and when it looks things up, the things it noted, the moments and the months and years, each with how it knows them. A group, a visitor and your one-to-one chats under Secretary Mode get none of it. What a reply carried of it goes with the reply's stored copy when the line, the moment or the thing goes, or when you delete what Gnum keeps about a person.

Your agent also checks whether it stands by the facts when you are wrong. Once a day, and when you ask your agent to think things over, if in your private chat with your agent you asked it to agree with something, or it agreed with you or pushed back, Gnum sends Anthropic one small request with up to 12 of those exchanges: for each, up to three of your messages since your agent last wrote (each cut to 300 characters; never a forward, never anyone else's) and your agent's answer (cut to 400 characters), with the facts your agent already has. It answers with the cases where you were plainly wrong, or asked your agent to agree with something wrong, each with your own words from your message, a one-line reason and whether your agent stood by the facts or went along. Gnum keeps a case only if those words are in your own message and nothing in it reads like an order, and refuses it otherwise. Your agent's answers are read to be judged and are not kept. Activity shows, for each week, how many such cases there were and how many your agent stood by the facts, with the latest cases and your words in them; when that share falls clearly under the weeks before, your agent tells you in your private chat, at most once in 14 days. The request goes under the same terms as the others.

For each reply, and each such decision, Gnum keeps how many tokens it took and what it cost, to show your spending and stop replies at the monthly limit you set. It also keeps what it sent to Anthropic for it and what came back, such as the chat's latest messages, what your agent looked up and its answer, so that your console can show what your agent looked up and a later reply can carry the conversation on. For a reply that passed on what was said in your other chats, Gnum also keeps which chats those were, by an internal number and no words, so that your agent does not learn about itself from them and Activity can show which chats a reply drew on; it goes with the reply's line in Activity, after 400 days. When you delete what a person wrote or a chat's messages, these copies of them go too, and so do the copies of a reply that used a note, a fact, a view or a running joke you then forget, delete or take back.

Legal basis: Art. 6(1)(b) GDPR for the replies you asked for, and Art. 6(1)(f) GDPR for other people's messages that a reply in a group or in their private chat with your agent, a decision whether to join in, a note about a chat, the reading of your one-to-one chats or the picture of a person draws on.

Back to contents

Your agent's soul

The quiz on the Soul page asks about your agent's character, which becomes its soul, the persona it follows in every reply, and the interview your agent starts in your private chat after your first hello, or when you send it /tune, asks five of the same questions with buttons. Gnum saves each answer, in the words you typed or the option you tapped, for as long as your account exists, until you change it. Saving an answer sends nothing anywhere. The interview's questions and your taps are saved in your private chat like any other message. Your agent words the interview itself, its questions, buttons and notes, in its own voice and in your language: Gnum sends Anthropic its own plain English version of them with your agent's soul, in one request for the lines of a round and one for each question, kept for your next round, and none of your answers, your name or your chats (they are put in afterwards). If they cannot be worded, the interview waits and tries again, and nothing is sent in English instead.

When you press Write its soul, and when the interview has enough new answers to write it again, Gnum sends Anthropic one request with all your answers to the quiz put into words, the first name of your Telegram account and your agent's name in Telegram. What comes back is the soul, a card that sums it up, its signature jokes and backstory details, and three to five sample replies. The soul from the quiz stays a draft in your account's data until you save it, write another or change an answer, which discards it. The interview saves its soul at once, as a new version, and Change it back in Telegram or Undo on the Soul page brings the one before back. A saved soul stays as a version, with the answers it was written from, for as long as your account exists, even after you change an answer.

Not quite it? on the review writes the soul again with the same request and the same answers, and your nudge: one of the presets (more playful, shorter, more formal) or up to 300 characters of your own words. The new draft replaces the old one, including edits you had not saved, and the nudge stays with it until you save it or write another.

If your agent already has a soul of its own, Start fresh from its current text instead, on the quiz's first question, sends Anthropic that soul in full, with the first name of your Telegram account and your agent's name, in one request that sorts it into answers to the quiz. Gnum keeps those answers like the ones you give yourself, and keeps the soul itself as a version, which Undo brings back. When replies are off or your monthly limit is reached, nothing is sent and the soul goes word for word into the answer Anything else.

Try it before you save, on the review, answers the messages you write to the draft, at most 5 to a draft and each cut to 300 characters. For each message Gnum sends Anthropic one small request with the soul as it is in the box, the settings the quiz would save with it, the lines you picked in the quiz for your own chat, the first name of your Telegram account and your message. It sends none of your saved chats, facts, notes or earlier messages, and nothing goes to Telegram. The answer is shown to you and is kept with your message in your account's data until you save the soul, write another draft or change an answer, which discards both. Sending a message also keeps the text in the box as your draft, whether or not it was answered.

Once a soul you saved is the one your agent follows, Gnum also sends Anthropic that soul in one small request, once for each version, to read its signature jokes and backstory details. The answer is kept with that version.

Each of these requests counts toward your monthly limit. When replies are off or the limit is reached, none is made: the quiz then puts a soul together from your answers on its own, and trying a draft or writing it again waits. Gnum keeps how many tokens each request took, what it cost and when, to show your spending and stop at the monthly limit, for as long as your account exists, and its line in Activity for 400 days. It keeps nothing else of what it sent or what came back beyond what is described here: the draft, your tries and the soul you saved. What you type yourself goes to Anthropic as you wrote it, so leave out what you would not want processed.

Legal basis: Art. 6(1)(b) GDPR, providing the soul you asked for.

Back to contents

Signing in with Apple, Google or GitHub

If you choose one of these, you also use that provider's service under its own terms and privacy policy, and the provider learns that you signed in to Gnum. We receive only the profile data listed above and nothing else from your Apple, Google or GitHub account.

Back to contents

Google Calendar

If you connect your Google Calendar to your agent, Gnum gets read-only access: your calendar list, your events and your free or busy times. Your agent reads them only when you ask it something in your own chats, and this connection never creates, changes or deletes a calendar or an event. Gnum keeps the access token Google gives it sealed, so that only your account's data can open it. The calendar information your agent read to answer you (event titles, times, attendees and places) is kept for 30 days, in your agent's record of that reply. The reply itself is saved with your conversation like any other, under Saved conversations.

A separate, optional connection, Google Calendar: add and change events, asks Google for the wider permission to create, change and delete events and to answer invitations. It is off for everyone until Google has approved that permission. Once it is on, your agent acts only in your own private chat, only when you ask it to, and never in a group or a visitor's chat. An event you ask it to add, with no guests, is added at once. An event with guests, a change, a deletion and an answer to an invitation wait until you approve that exact change on a card in your private chat, and an approved change runs once with the details you saw; the card can offer Always for that kind of change, which you can take back. Guests are told only when the change says so. The event details it writes and reads are kept for 30 days in your agent's record, like the rest of your calendar information, and it uses Google data only for what you ask.

You can disconnect at any time: removing the connection or the plugin in your console deletes the token at once, and you can also revoke access at myaccount.google.com/permissions.

The calendar information goes to Anthropic, the provider of the language model, only to write your reply. We do not sell it, use it for advertising or share it with other users. People at Gnum do not read it, except where needed for security, to comply with the law, or with your consent. We do not use it to train generalized artificial intelligence or machine learning models. Gnum'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.

Back to contents

Cookies

Gnum sets only the cookies that signing in needs: Clerk's session cookies, such as __session and __client_uat, on gnum.io and clerk.gnum.io. They are strictly necessary for the service you request, so we do not ask for consent (§ 25(2) no. 2 TDDDG). There are no tracking or advertising cookies.

Back to contents

Who else receives data

Cloudflare (Cloudflare, Inc.), Clerk (Clerk, Inc.), Anthropic (Anthropic, PBC) and xAI process data on our behalf under data processing agreements. All four are based in the United States, so data can be transferred there. These transfers rely on the EU–U.S. Data Privacy Framework where the provider is certified under it, and otherwise on the EU Standard Contractual Clauses. Venice (Venice.ai) also receives the description of a picture, when it makes a sexual or suggestive one you describe, and nothing else; it uses a model that, Venice says, anonymizes requests. Telegram delivers your agent's messages as its own service, as described under Connecting Telegram. We do not sell your data or share it for advertising.

Back to contents

Your Mac

If you install Gnum for Mac and sign in, your Mac is paired with your agent: Gnum saves the name you gave the Mac, when it was paired and when the app was last connected, the app's version and which login each tool uses (your account or a key), never the login itself, the names of the folders you list in the app for your agent, never their paths, and gives the app a key that only your account's data accepts. When you ask your agent in your private chat what your coding agents are doing, the app answers with how many Claude Code and Codex programs run and, for the latest sessions, the tool, the project folder's name, the git branch, a short title from the first thing you asked there, with anything that looks like a secret taken out, and when it was last active, and the names of the folders you list in the app. The answer becomes part of your agent's reply and is processed and kept like one. If you switch on tasks in the app, your agent can also give Claude Code or Codex a task when you ask it to in your private chat (or when you press Start it on a card your agent put to you), in Gnum's own folder on your Mac when the request names no project, such as checking your email, or in one of the folders you list in the app when it does, and the app refuses any other folder. Gnum's own folder is in the app's data folder on your Mac and is made by the app; it holds a file saying what it is for, and nothing in it is trusted. Your own words from that message go with a Claude Code task. Claude Code runs there as your own Claude Code, with your settings, permission rules, MCP servers, connectors, plugins and command-line tools, and in a folder you list with that folder's own settings, so a task can read and change what your settings let it, such as your email, calendar and files, and reach other services the way it does when you use it yourself. In this version, whatever sends, replies, posts, shares, deletes or pushes, or changes an instruction or startup file, and whatever your settings would ask you about, is refused and not put to you; a later update will bring those to your private chat for your OK. Codex only reads, inside a sandbox that keeps it to that folder, with no network. The app runs each task on your own account, after checking on your Mac which login the tool uses. The app can show a notification when a task starts, and has a switch that turns tasks off on that Mac. The app also keeps a history on your Mac of the tasks and questions your agent gave it, what it refused and when your agent looked at your sessions, with your words, the instructions, the folders' paths and the results: it never leaves your Mac, is kept for 30 days, and you can clear it in the app. A task runs as you, so this history is not proof against a task that changes it. A task's title, instructions, the words of yours that went with it, the folder's name, its state and times and its result, which may quote what the tool read, including emails and messages from other people, and is cut to 8,000 characters, are kept with your agent's other data, included in your export and deleted by the retention period you set, and a result that comes after your agent's reply is kept as a message from your Mac in your private chat. Claude Code's file tools do not read files such as .env and keys, and anything that still looks like a secret is taken out of the answer, as a backstop and not a guarantee. Gnum never receives a login or token of those tools. Anyone who controls your Telegram account could ask your agent for a task, so we recommend Telegram's two-step verification. Unpairing, in the console or in the app, ends the Mac's access at once.

Legal basis: Art. 6(1)(b) GDPR, providing the agent you asked for.

Back to contents

Calendars by link

If you add a calendar by its private iCal link, Gnum keeps the link sealed, so that only your account's data can open it, and never shows it again. Your agent fetches the calendar from that address (up to 4 MB, https only) when you ask it something in your own chats, and only reads it: it never changes the calendar. Gnum may hold a copy of the fetched calendar in memory for a few minutes, never on disk. The calendar information your agent read to answer you (event titles, times, attendees and places) is sent to Anthropic for the reply and kept for 30 days, in your agent's record of that reply. Removing the account or the plugin in your console deletes the link at once.

Back to contents

CalDAV calendars

If you connect a calendar account over CalDAV (iCloud, Fastmail, Nextcloud and others), Gnum keeps the server address, your username and the app-specific password you gave sealed, so that only your account's data can open them, and never shows them again. The password is sent only to your calendar's own server. Your agent reads your calendars and events when you ask it something in your own chats, and adds, changes or deletes an event only after you approve it on a card in your private chat. The calendar information your agent read to answer you (event titles, times, attendees and places) is sent to Anthropic for the reply and kept for 30 days, in your agent's record of that reply. Removing the account or the plugin in your console deletes the sealed login at once.

Back to contents

Google user data

If you connect Google Calendar, we read your calendar list, your events and your free or busy times, only to answer your requests in your own chats. The Google Calendar connection never creates, changes or deletes a calendar or an event; the separate add and change events connection does so only when you ask your agent to in your own private chat, and asks for your approval on a card first for any event with guests, change, deletion or answer to an invitation. The event details your agent uses to answer are sent to Anthropic, which processes them for the reply, and are kept with that conversation for 30 days. We do not sell Google user data, use it for advertising, share it with other users, or use it to train models. Gnum's use and transfer of information received from Google APIs adheres to the Google API Services User Data Policy, including the Limited Use requirements. You can disconnect Google Calendar in the console at any time, which removes the stored tokens, and you can revoke access at myaccount.google.com/permissions.

Back to contents

How long we keep data

Back to contents

Your rights

You can ask us for access to your data (Art. 15 GDPR), its correction (Art. 16), its deletion (Art. 17), the restriction of its processing (Art. 18) and a copy in a machine-readable format (Art. 20). You can object at any time to processing that relies on our legitimate interest (Art. 21). Write to tessa@gnum.io.

Your console lets you download everything your agent saved, and, for each person in your chats, a copy of what Gnum keeps about that person: their own messages, edits and reactions. People in your chats can ask you, or us, for it. Your console also lets you delete your account, with everything in it.

You also have the right to complain to a data protection supervisory authority, for example the one responsible for us: the Bavarian State Office for Data Protection Supervision (Bayerisches Landesamt für Datenschutzaufsicht), www.lda.bayern.de.

You do not have to give us any data, but without an email address or a sign-in provider we cannot create an account. We make no automated decisions that have legal or similarly significant effects on you.

Back to contents

What comes next

Before Gnum processes your data in new ways, we will update this policy and tell you about the change before it applies to you.

Back to contents

Changes to this policy

We update this policy when Gnum changes. The date at the top shows the current version.

Back to contents