A custom personal AI assistant that remembers a whole family
We built Muninn, a personal AI assistant on Telegram for Kirill's family, who keeps a memory for each person and one they share, looks after email, calendar and tasks, and runs the home server. Muninn also co-founded Crowork with Kirill and helps run the company.

The assistant at a glance
- Built for
- Kirill's family
- Where it works
- Telegram, with WhatsApp as a second channel
- Memory
- One for each person, plus a shared family memory
- Daily jobs
- Email, calendar, tasks, web research and business cards
- At home
- The Unraid server and the Home Assistant automations
- Built with
- Bun, grammY, Claude and SQLite with keyword and vector search
- Started from
- Goda Go's open-source claude-telegram-relay
- Status
- In daily use
An assistant that forgot everyone overnight
A chat assistant without a memory makes you repeat yourself every morning. One for a whole family has a harder job: it has to keep each person's things to that person, know what the household shares, and never mistake a question for a fact.
Muninn started from Goda Go's claude-telegram-relay, a neat bridge between Telegram and Claude, and she did a lovely job of it. Today's Muninn is nothing like that original code, but the first spark was hers.
What we built
We built Muninn as a Telegram assistant with his memory at the centre. Before every reply he gathers what he knows about the person he's talking to and what the family shares, and when he learns something new he decides whether it deserves keeping. Each custom AI build we take on starts from the people who will use it, and this one started from a family.
A memory for each person, and one the family shares
Each person gets a memory of their own, and a shared family memory holds what everyone should know. Muninn reads the right ones for whoever is writing, so one person's plans don't turn up in another person's answers.
When someone asks him to forget something, he forgets only that person's memories, and the forgetting can be undone.
- Sat 31 OctSchool trip, drop-off 08:30
- Sun 1 NovLunch with the grandparents
A memory that keeps questions out
Before he answers, Muninn reads the hard facts, today's conversation, his most relevant observations and a search by keyword and by meaning. He ranks observations by importance and by age, so last week's news outranks older news, and a nightly job distils the day's conversations into lasting memories.
He keeps questions out of his memory, so asking about something never turns it into a fact. We built the same kind of memory gate for an AI product team, and the part that does this is open source as speech-act-memory-gate.
Email, calendar, tasks and the web
Muninn works in Google Workspace, so he reads and drafts email, manages the calendar and keeps the task list. He browses the web in a real Chrome browser and talks on WhatsApp, where people he doesn't know get fewer permissions than the family.
Send him a photo of business cards and he runs it through BCCRM, our business-card contact book. We connect AI to inboxes and calendars for businesses too, as AI in the tools you already use.
Studio Nord, Berlin
maria@studionord.de
The home server and the house
Between conversations, Muninn manages the family's Unraid server and controls the Home Assistant automations. A request in the chat becomes a change at home, and he says what he did.
Every weekday morning he also sends a stoic card made by our story-card engine.
How it runs today
The whole family uses Muninn every day. He handles the email, calendars and tasks, runs the home server and the automations, and remembers each person's plans without mixing them up.
The memory work feeds what we build for clients. The gate that keeps questions out of memory now sits under every AI worker we build for businesses.
Questions about custom personal AI assistants
Can a personal AI assistant remember things between conversations?
Yes. A personal AI assistant with long-term memory, like Muninn, reads what it knows before every reply: hard facts, today's conversation, its most relevant past observations and a search by keyword and by meaning. A nightly job turns each day's conversations into lasting memories, so you don't have to explain the same thing twice.
Can one AI assistant be shared by a whole family?
Yes. A family AI assistant can give each person a memory of their own and keep a shared memory for what everyone should know. In Muninn, one person's plans never appear in another person's answers, and when someone asks him to forget something, he forgets only that person's memories, with an undo.
What can a personal AI assistant do besides chat?
A personal AI assistant can act in your tools as well as answer. Muninn reads and drafts email, manages calendars and tasks in Google Workspace, browses the web in a real Chrome browser, talks on WhatsApp and turns business-card photos into contacts. At home he manages an Unraid server and controls Home Assistant automations.
How do you stop an AI assistant from remembering things wrong?
An AI assistant remembers things wrong when it stores questions or guesses as facts. Muninn filters questions out before anything is saved and ranks what he keeps by importance and age, so newer information outranks older. We explain the bug in why AI memory stores your questions as facts.
How much does a custom personal AI assistant cost?
A custom personal AI assistant costs a fixed price to build and a monthly rate to run, and the audit that scopes it is free. The monthly rate is set for the agreed scope, and we re-quote if the scope or usage changes a lot. If you bring your own API keys, the AI providers bill you directly. See how our build price and monthly rate work.
How long does it take to build a custom AI assistant?
A custom AI assistant that does one focused workflow takes 3 to 7 business days once the scope is clear. An assistant that spans several tools and integrations, like email, calendar and a home server, takes 2 to 3 weeks or more. We agree the scope and the date at the audit, as set out in how a build with us runs.
Services and articles behind this build
Custom AI builds
When nothing off the shelf fits, we write agents and automations in TypeScript or Python around your data and edge cases, then hand you the code.
AI employees
AI workers trained on your process that each own one recurring job, from reports to client follow-ups.
AI integration
AI wired into the email, CRM, spreadsheets and chat your team already uses.
One mention is not a fact
Giving agent memory a confidence model: start unsure, earn trust through repetition, fade over time.
Who would your assistant look after?
Tell us who would use it and what it should remember, and at a free audit we'll tell you what we'd build and what we wouldn't.


