Asking Pepper
You can ask Pepper anything about the chats you connect and the things it tracks for you, in plain words. It answers briefly, with its sources.
- Telegram. Send a message to your Pepper bot. A voice note works too: Pepper shows what it heard, then answers. By default it answers a voice note with a voice note. Send a photo, such as a screenshot of a chat, and the caption is your question.
- The dashboard. Type in the Ask Pepper anything… bar at the foot of any tab. Press
/to jump to it. - By phone. Call Pepper from your own number and say your passphrase. Pepper is asked to answer in at most three short spoken sentences, with no links or source lines. See Talking to Pepper by phone.
A conversation carries on for 45 minutes after your last message (the threadMinutes setting), so follow-ups
like "and what did she say after that?" work. Telegram and the dashboard each keep their own conversation, and
every phone call starts a new one. Send /new on Telegram to start fresh there.
On Telegram and by phone, not every message is a question. Something that reads like a to-do, like remind me to book the flights Friday,
becomes a task instead. See Follow-ups.
By default, Pepper answers only from what is already stored on the Mac. It finds answers with a fixed set of lookups, which read your records and never change anything:
| Lookup | What it gives back |
|---|---|
| Search messages | Matches by words, and by meaning when that's turned on, filtered by person, chat, dates or "from me" |
| Read a thread | One chat in order, for context |
| List chats | Chat names, unread counts, last activity |
| List people | Who you talk to, VIP or not (People), how often, last contact |
| Open items | What you owe, what's owed to you, and your task list |
| One person | Facts Pepper has kept about them, what you owe each other, recent decisions, last contact |
| Decisions | What was agreed, by whom, and when |
| Priorities | Your priorities, in your order |
| Exact meeting quotes | Off by default. Quotes from one meeting, found by Tactiq, the meeting transcription service, which receives your question. See Tactiq meetings |
Each question has limits: 12 lookups by default (twice the askMaxIterations setting), 20 messages per
search, 40 per thread, about 300 characters of each message, and about 6,000 characters per result. Ask
something narrower if an answer feels thin.
Chats you set to ignore with /ignore are left out when Pepper searches messages, reads a thread or lists
chats. That filter doesn't cover exact meeting quotes or the people lookups.
Pepper cites the messages it used, by sender, chat and time, and quotes a few words when the wording matters.
When an answer has verified sources, a final sources: line on Telegram lists up to four of them by sender and
chat, like Dana Reyes, Lease 2026. On the dashboard they appear as tags under the answer.
Every message in that list is checked against what the lookups actually returned. If the AI model behind Pepper names a message it was never shown, that source is dropped and the answer says how many were dropped (on Telegram, only when at least one source remains). The check covers the source list, not the wording of the answer itself.
Pepper is told to say so when its records don't answer the question, and never to guess what someone said or make up a date, amount or name. It can still get things wrong, so check anything that matters against the message.
A link only appears in an answer if it was in a message Pepper was shown, character for character. Anything
else that looks like a link becomes [link removed]. So a message written to trick Pepper has a hard time
slipping a fake link into an answer you trust. The filter matches the common forms a link takes, and a link
copied from a message is only as safe as that message. Telegram link previews are off too.
If you need a link, ask for the message itself, like "what link did Mira send about the venue?".
One-time codes, verification codes and login links are masked in the text Pepper reads, and again in what it
sends you. A message that says Your code is 482913 reaches the model as Your code is ••••••. Links that
carry a secret, such as a password reset link, become [secret link removed]. The masking looks for the
common forms a code takes. A photo is not masked: the model is told not to repeat a code it can see in one,
which is a weaker guard.
So Pepper won't read you a code, even if you ask, but don't count on that. Open the app the code was sent to.
What did Amber Liu say about the budget last week?
What do I owe Theo Grant?
Who am I waiting on?
When did Lena Park and I agree on the launch date?
Summarize the Q4 planning group since Monday.
What's Mira Chen's role?
Draft a reply to Sam Lee saying Thursday works.Names, dates and chat names help most. "Last week" and "in September" are turned into date filters for you.
- It can't send, edit or delete anything, in any chat. Ask for a reply and you get a draft that you copy and send.
- It can't browse the web, open files or run anything.
- It can't see your calendar yet.
- It can't read PDFs, videos or stickers. It reads text, voice notes and images up to 5 MB.
- It won't take orders from messages. Text other people wrote is labelled as content to report, and Pepper is told never to treat it as an instruction. That cuts the risk without removing it; see How Pepper keeps your data safe. A forwarded message is summarised, never turned into a task.
Your question, any photo you attach, the conversation so far and the records Pepper finds are sent to Anthropic, whose model writes the answer, and Anthropic's retention terms apply. A voice note's audio goes to OpenAI if you chose OpenAI for transcription. Details are in How Pepper keeps your data safe.