Companion at work and at home
Companion Memory keeps your messages, calendar, photos and documents on a box you own, and it learns from them there. It learns how your weeks usually go, finds to-dos in your messages, and updates what it knows each night. This page has 16 examples of what that looks like, eight at work and eight at home. Some of them include a short Behind the scenes part that names the jobs, models and containers involved. For how the learning works, see How Companion learns.
Each example marks its features with the four status labels defined on the personas page. In short, Works today is on by default, Behind a setting stays off until an owner or operator turns it on, Not yet is not built or not connected, and Planned is designed with no code yet.
The people on this page are fictional. Most come from the personas page, and four are described in a line instead. None of them are real users. Nothing here is a testimonial.
This page describes Companion Memory main and Companion Hub dev in November 2026. Clock times on this page are in UTC, because the Memory containers set no time zone unless the operator sets one. The latest release, Memory v2026.11.11, predates some of it: there, activity types come from a keyword labeler instead of a language model.
At work
Eight examples from six people whose weeks run on meetings, messages, receipts and deadlines.
Noor gets ready for client meetings
The situation. Noor is an independent consultant who runs client meetings. She prepares agendas, invites clients and keeps minutes. This week she has three client meetings, and one client’s email thread asks for a revised proposal before Thursday.
What Companion does today.
- Works today: her calendar comes in from Outlook or Google, and each meeting gets a plan with the agenda, the attendee list and her notes in one place. The Google import reads the primary calendar only.
- Works today: Companion reads her new email on her box and turns requests into to-dos, such as “Send the revised proposal”. Her Autonomy setting decides whether a to-do joins her list or waits for her. See From a message to your to-do list.
- Works today: the Calendar shows what she usually does at each time of the coming week, learned from her own history. She can see when she usually does client prep before she books another call.
Behind the scenes. Every 30 minutes a local language model scans her new mail for requests. By default it is the same gemma4:e4b model that answers chat. When she asks Chat about a client, it searches her mail, calendar and people and links each item it used. An hourly check looks at tasks whose calendar time has passed for signs they got done, such as a message that says she finished it, and on Balanced it asks her before marking one done.
What’s coming.
- Planned: spotting commitments in conversations, such as “we’ll have the draft to you Friday”, and staging a calendar event or task for her to confirm.
- Not yet: small decision models trained on her own examples, to answer “is this a real commitment?” quickly on her box. Nothing is trained yet.
- Not yet: separate matters, so each client’s material stays apart. Search and grants apply to the whole account today.
Try it.
- Open Planning, then Tasks.
- Look in the Proposed column for to-dos that are waiting for you.
- On each card, choose Plan to keep it without a time, Schedule to give it a time, or Dismiss to drop it.
Sam Okafor follows up after an event
The situation. Sam is 45 and works as an event photographer. After an event, attendees write to ask whether he has photos of them. He answers from his archive and sends only what each person is entitled to receive. He also owes the organizer a set of photos by the end of the week.
What Companion does today.
- Works today: each request that reaches him through Companion waits in his review inbox, and he sees every answer before it is sent. He can answer with less than was asked.
- Works today: requests in his email and texts turn into to-dos, so “Send the event photos to the organizer” is waiting for him on Monday morning.
- Works today: his archive and everything Companion learns from it stay on his box. Nothing is sent to a cloud service to be learned from.
What’s coming.
- Planned: a personal tone add-on that learns his writing style, so replies Companion drafts for attendees sound like him. He would still review each one before it is sent.
- Planned: duplicate-contact detection, so an attendee who writes from two addresses shows up as one person.
- Not yet: a rule for photo deliveries to attendees, and priced requests with payment for a delivery.
Sam Okafor gathers his receipts at tax time
The situation. Sam’s gear, travel and software receipts are spread across email, PayPal, Amazon and ride apps. His accountant wants them sorted before the deadline.
What Companion does today.
- Works today: he imports his Amazon orders, PayPal history and an archive of his mail. Amazon orders and receipts found in his mail become receipts with their line items, under Memory, then Receipts. A PayPal payment becomes a receipt with the merchant and total only. Uber and Lyft rides import as trips with their fares, and Uber Eats orders become receipts. Order emails that the mail import would drop as bulk mail are kept when they look like a receipt.
- Works today: he filters his receipts by date and saves the search, so he can run it again next year.
- Works today: each physical item on a receipt becomes a Thing, so he can see which camera body came from which order. Details such as a serial number stay on the Thing.
- Works today: each day’s spending on physical items, the Things from his receipts, goes into the same daily numbers as everything else, and Companion flags days that stand out from his usual spending.
Receipts and totals stay on his box, and nothing is sent to an accounting service.
What’s coming.
- Not yet: receipts from live Gmail and Outlook sync. Live sync keeps order emails as messages only, which is why Sam imports his mail archive.
Dr. Amara Osei protects her writing time
The situation. Dr. Amara Osei is 52 and a sleep researcher. She is recruiting volunteers for a study, and their questions and scheduling requests arrive all week. She also needs long, quiet blocks to write up her work.
What Companion does today.
- Works today: messages from volunteers that ask for something, such as moving a session, turn into to-dos, so none of them gets lost in a busy inbox.
- Works today: the Calendar shows what she usually does at each time of the coming week. Before she offers a volunteer a slot, she can see which hours usually go to her own work.
What’s coming.
- Planned: proactive insights that interrupt only when it’s worth it, such as a note when a new booking would land on time she usually keeps for writing.
- Planned: consent receipts that show the consent basis for each disclosure, which she could show to an ethics board.
- Not yet: priced requests, and a way to pay volunteers.
Try it.
- Open Planning, then Calendar.
- Select an empty time in the next seven days. An Add to calendar panel opens.
- Read Predictions to see what you usually do at that time. Under Possibilities, pick a to-do to put there.
Dr. Amara Osei hands a research errand to an agent
The situation. Amara has a task to compare actigraphy devices for a new study. She wants an agent to do the legwork while she keeps control of what it touches.
What Companion does today.
- Works today: with OpenClaw installed on her Hub, she opens the task in Planning and chooses Dispatch to OpenClaw. A chat thread shows the job’s progress. She can let it continue, redirect it with a message, or halt it, and the result comes back to the task.
- Works today: when OpenClaw calls a Memory tool that sends, deletes or reaches an outside service, the call waits in Approvals, whatever her Autonomy setting is.
- Works today: the agent runs in a container on her Hub and uses the Hub’s local model. It reaches the web or a cloud model only if she turns that on in OpenClaw or adds a cloud provider key on the Hub.
Behind the scenes. Memory puts the job on its agent queue and asks the Hub to wake the OpenClaw container. OpenClaw fetches the job from Memory, gets its model settings from the Hub, and posts checkpoints and the result back to Memory. More in OpenClaw.
sequenceDiagram
accTitle: How a task dispatched to OpenClaw runs on the Hub
participant A as Amara
participant M as Memory
participant H as Hub
participant O as OpenClaw container
A->>M: Dispatch to OpenClaw
M->>M: Put the job on the agent queue
M->>H: Ask the Hub to wake OpenClaw
H->>O: Wake OpenClaw with the job id
O->>M: Fetch the job
O->>H: Work with the Hub's local model
O->>M: Post checkpoints
A->>M: Continue, redirect or halt
O->>M: Read her steering, post the result
M->>A: Result on the taskAn engineering manager wins back focus time
The situation. An engineering manager notices too late that meetings have taken over the week. The blocks they keep for deep work keep getting booked over.
What Companion does today.
- Works today: calendar sync from Google or Outlook adds up meeting time for each day. When they ask Chat “anything unusual this month?”, it says that meeting time is well above their usual level, and it shows how many days the finding rests on.
- Works today: under Self, then Profile, they type a task such as “prepare the quarterly review” and choose How I’d do this. Companion writes steps in the way they usually work, from what their memory shows. Adding the steps puts them under Possibilities on the Calendar and books nothing.
- Works today: they ask Chat to schedule a focus block. With no time named, Companion puts it in the first free gap within working hours that fits, and it never moves booked time. On Balanced, the booking waits in Approvals only when it follows a lookup in the same chat turn.
What’s coming.
- Planned: proactive insights that interrupt only when it’s worth it, such as a note when meetings start to crowd out focus time.
Priya follows up after a conference
The situation. Priya is 33 and new in town, and she is building a social circle. She is careful about who she trusts. After a conference she has new contacts, a few invitations to coffee, and some requests to follow up on.
What Companion does today.
- Works today: requests in her new email and chats, such as “could you send me the slides?”, turn into to-dos. Companion reads them on her box.
- Works today: when another Companion asks her something, she decides each request before anything is sent.
What’s coming.
- Planned: spotting commitments in conversations, such as “coffee on Thursday at 10?”, and staging a calendar event for her to confirm.
- Planned: duplicate-contact detection, so a new contact who writes from a work address and a personal one shows up once.
- Not yet: confirming a claim such as “we met at the conference” with a verifiable credential.
A video editor keeps their AI responsive in a deadline week
The situation. A video editor has a workstation and a mini PC, and each runs its own Hub. During renders the workstation’s graphics card is busy, so chat and background work slow down.
What Companion does today.
- Works today: Hub Pool, which is on by default, lets the two Hubs share models once the editor pairs them over their own Tailscale network. Each model request goes to the Hub that has the model and the shortest queue, and to the other Hub if one stops answering.
- Works today: Pause background work on the Jobs page stops every queue that no one is waiting on. Chat keeps answering, and nothing is lost: jobs keep queuing and run once the switch is off again.
Model requests move only between Hubs the editor runs. Nothing goes to a cloud provider unless they add a provider key. More in Pool two Hubs and Hub Pool.
Try it.
- Open Jobs.
- Open the Controls card.
- Turn on Pause background work. Turn it off when the deadline has passed, and the queued jobs run.
At home and in daily life
Eight examples covering family logistics, film night, routines, sleep, errands, the things you own, photos and caring for someone.
Maya keeps the family week on track
The situation. Maya is 38 and the household operator for her family. She bought a family box for photo memory and privacy. School emails, appointment confirmations and group chats all land with her, and when something needs sorting out, she is the person others ask.
What Companion does today.
- Works today: requests in those messages, such as “please return the signed form by Friday”, turn into to-dos. In the default Balanced setting, a to-do Companion is sure about joins her list, and one it is less sure about waits in Proposed.
- Works today: the Calendar shows what usually fills each part of her week, so she can see where an errand is likely to fit.
- Behind a setting: web push reminders in a browser. Push needs keys set on the box.
Behind the scenes. Each night at 02:30, after a local language model gives each new calendar entry a type such as errand, her box rebuilds a statistical profile of how often each routine, such as a haircut, comes round. The forecast on her Calendar draws on it. Turning overdue or soon-due routines into suggested time slots is Behind a setting: it runs only with smart rotation on, and no screen in the web app shows those suggestions yet.
What’s coming.
- Planned: proactive insights that interrupt only when it’s worth it, such as a heads-up when two family plans clash.
- Planned: filtering noise before it is stored, so newsletters and promotions take up less of the family’s memory.
- Not yet: push notifications in the phone apps.
Try it.
- Open Settings, then Autonomy.
- Choose Cautious if you want every to-do to wait for you, or keep Balanced.
- Select Save.
Maya picks a film the whole family will like
The situation. Friday film night at Maya’s house used to start with a long scroll, because nobody remembered who liked what. Her family’s tastes are already in her memory, from imports such as Netflix, Letterboxd, Spotify and Goodreads.
What Companion does today.
- Works today: she asks Chat for “something sci-fi tonight that my husband and the kids would all enjoy”. Companion finds her family in her memory and reads each person’s likes and dislikes. The chat model runs on her Hub.
- Behind a setting: looking up what is streaming, which needs a key for The Movie Database that the box’s operator sets. The lookup reaches outside her box. On Balanced, it runs at once when she asks for it directly. Here it follows the lookups of her family in the same chat turn, so it waits in Approvals until she approves it. Who likes what never leaves the box: the movie database receives only a genre and a content type.
What’s coming.
- Not yet: the streaming lookup on a box installed through Companion Hub. The Hub’s install settings have no field for the key, so the lookup reports an error.
- ChatMaya asks for a sci-fi film for the familyHer phone
- Chat modelPlans which tools to callHub
- people.relationshipsFinds her husband and kids in her memoryMemory
- media.entertainmentPreferencesReads each person's likes and dislikesMemory
- ApprovalsMaya approves the outside lookupHer phone
- media.streamingAvailabilityAsks the movie database for sci-fi titles and servicesOutside
- AnswerTitles that suit everyone, and where to watch themHer phone
Devon gets to know his own routine
The situation. Devon is 29 and an early adopter. He runs his own box on a network-attached storage device, and he reads technical specifications closely. He wants to see how his weeks really go: when he works, when he exercises and when he sleeps.
What Companion does today.
- Works today: Companion sorts each day of his timeline into broad activities, such as work, exercise, social time, errands, leisure, travel and sleep. A language model on his box does the sorting.
- Works today: a small personal model learns his routine from that timeline. Once it has learned a pattern, the Calendar shows what he usually does at each time of the coming week.
- Works today: each night, Companion updates a small personal add-on from the previous day’s events. Since Companion Memory 2026.11.11, each update is saved as a new version and the three before it are kept, so an update can be rolled back by hand. There is no button for that yet. An update that goes wrong leaves the last good version in place.
- Behind a setting: the Forecast page, a fuller view of tomorrow’s forecast. It is off until the box’s owner turns it on, and Devon runs his own box.
- Works today: for apps and agents he connects, a quick-decision service on his box answers yes-or-no and pick-one questions by comparing meaning. When it is unsure it falls back to a simple word-overlap check; passing those questions to the local language model is not wired yet. No built-in feature uses it on its own yet.
Companion shows patterns from his own timeline. It does not give health advice.
What’s coming.
- Not yet: small decision models trained on his own examples, to answer questions like “is this worth remembering?” on his box. The trainer is not finished, and nothing is trained yet.
- Not yet: training on another device he owns, such as a desktop with a bigger graphics card, through Companion Hub.
- Planned: quality checks that every newly trained model must pass before it is used, on accuracy and on how well its confidence matches reality.
A new parent looks at what affects their sleep
The situation. A new parent feels worn out and suspects late-night browsing is part of it, but cannot tell from memory what changed and when. They import an Apple Health export and install the browser extension.
What Companion does today.
- Works today: Apple Health sleep records become sleep events. Screen time counts only the browsing the extension reports, not all phone use.
- Works today: every hour, Companion rolls each day into one row of numbers, such as sleep, screen time, meeting time, messages and spending. Plain statistics then look for numbers that move together and for days that stand out.
- Works today: they can ask Chat “anything unusual lately?” or “does screen time go with my sleep?”. The answer shows its evidence: how many days, how strong the link is, and which sources. Self, then Stats, shows the daily numbers.
Companion reports numbers that move together. It does not claim that one causes the other, and it gives no health advice.
Behind the scenes. The rollup and three statistics engines run in the Memory api container with no language model and no outside call. The chat model only puts their findings into words.
Tomas handles the group’s errands
The situation. Tomas is in his thirties or forties. He organizes a standing group of five friends from three households, plans dinners and films, and buys tickets when the group agrees. The planning happens in group chats, and the errands land on him.
What Companion does today.
- Works today: a group plan keeps the details of each dinner or film in one place.
- Works today: requests in the group’s messages, such as “can someone book the table?”, turn into to-dos for him.
- Works today: the Calendar shows when errands usually fit into his week.
- Works today: anything that sends, deletes or changes something outside the box waits for his approval first. On Balanced, a lookup he asks for directly runs at once.
- Behind a setting: a free-or-busy check with each friend’s box. The answer shows free or busy only, with no event details.
Behind the scenes. His texts reach the box through BlueBubbles, which relays iMessage from his own Mac. Every 30 minutes a local language model, by default the same gemma4:e4b that answers chat, scans new messages for requests. If he asks Chat to reply “I’m in” to a friend, the reply waits in Approvals and leaves through his own Mac only after he approves it.
What’s coming.
- Planned: spotting when the group settles on a time, such as “Friday at 7 works for everyone”, and staging a calendar event for him to confirm.
- Not yet: buying tickets with an agent, and a spending limit for purchases.
A homeowner keeps track of what they own
The situation. The dishwasher breaks, and its owner cannot remember when or where they bought it. They also keep losing the spare car key.
What Companion does today.
- Works today: importing a mail archive or a shop’s order export builds receipts with their line items, under Memory, then Receipts. Order emails the import would drop as bulk mail are kept when they look like a receipt.
- Works today: each physical item on a receipt becomes a Thing, under Memory, then Things. An ebook does not. Importing the same mail again makes no duplicate, and Companion suggests merging Things that look the same.
- Works today: a room scan from Companion Scan, once sent to Memory, files each object it saw as a sighting. They confirm sightings on Things, such as “car key, kitchen drawer”.
Receipts, Things and sightings stay on the box. Memory receives a scan only when Companion Scan sends a finished one, and the app cannot read anything back from the box.
What’s coming.
- Not yet: receipts from live Gmail and Outlook sync, which keeps order emails as messages only.
- Not yet: warranty reminders. Things have a warranty date, but no screen shows it.
Ruth keeps fifty years of family memories
The situation. Ruth is 71, a grandmother and the family archivist. She keeps about fifty years of photos, letters and recipes, and she worries about other people seeing her things. She uses a voice assistant and a tablet.
What Companion does today.
- Works today: her archive sits on her family’s box, and Companion learns from it there. Nothing is sent to a cloud service to be learned from.
- Works today: she can search her photos, letters and recipes together from her tablet.
- Works today: export and import move her archive to a new box when she has one. Import replaces the data already on the receiving box.
Behind the scenes. Imports keep each photo’s time and location, and a local language model writes daily and monthly summaries. Under Self, then Reflections, her companion asks about her life and works out answers from her memory; on Balanced it saves the ones it is sure about and asks her about the rest. Place names come from an outside service unless she turns off Place lookups under Settings, then Privacy.
What’s coming.
- Planned: smarter search ranking, so the photo or recipe she means comes nearer the top.
- Planned: duplicate-contact detection, so a relative who appears under two names over the years can be recognized as one person.
- Planned: a personal tone add-on, so a note Companion drafts for her sounds like her.
Maya helps Ruth, on Ruth’s terms
The situation. Ruth lets Maya see her appointments and reminders, within limits Ruth sets and until a date Ruth chooses. Ruth can end it at any time. A permission does not replace legal authority, and Companion gives no medical advice.
What Companion does today.
- Works today: Ruth gives Maya a standing permission that names the kinds of question it covers. Other questions wait for Ruth.
- Works today: Ruth can revoke the permission whenever she likes, and new asks from Maya stop.
- Works today: when Ruth messages Maya to ask for a lift, the request can turn into a to-do on Maya’s list.
What’s coming.
- Planned: a care-circle digest for an older relative. It would show no location and would end on a set date.
- Planned: a ready-made care template. Today Ruth sets each limit herself.
A day with Companion
This timeline follows Noor through a working day, and it marks the steps that are still coming.
timeline
title A working day for Noor
Morning : She reviews to-dos Companion found in her email
Midday : The Calendar shows when she usually does client prep
Afternoon : Times agreed in a meeting wait as events to confirm (coming) : Follow-up drafts that sound like her (coming)
Night : Companion updates its picture of her weekIn the morning, Noor opens Proposed and plans or dismisses the to-dos Companion found in her email overnight. At midday, the Calendar shows when she usually does client prep, so she books her next call around it. Both of these work today. In the afternoon, two planned features would help: staging a calendar event when a meeting settles on a time, and drafting follow-ups in her own writing style. At night, Companion updates its picture of her week from the day’s events, which works today.
From a message to your to-do list
Works today: Companion reads new email, texts and chats on your box through the day. Each morning it also looks through the rest of your memory, such as calendar entries, for anything still left to do. It skips to-dos you already have.
What happens next depends on your Autonomy setting. You pick it during setup, and you can change it in Settings under Autonomy.
- Cautious: every to-do waits in Proposed until you choose Plan, Schedule or Dismiss.
- Balanced, the default: a to-do Companion is sure about joins your task list straight away. One it is less sure about waits in Proposed. Giving a new to-do a time on your calendar still waits for you.
- Hands-off: Companion adds the to-dos it is sure about and places them on your calendar on its own.
This diagram shows the path a new message takes in the Cautious and Balanced settings.
flowchart TB
accTitle: How a new message becomes a to-do in the Cautious and Balanced settings
A["A new email, text or chat arrives"] --> B["Companion reads it on your box"]
B --> C{"Is there something left to do?"}
C -->|No| D["Nothing is added"]
C -->|Yes| E{"Your Autonomy setting, and how sure Companion is"}
E -->|"Cautious, or Balanced and less sure"| F["Waits in Proposed"]
E -->|"Balanced and sure"| G["Joins your task list"]
F --> H["You choose Plan, Schedule or Dismiss"]
G --> I["A time on your calendar still waits for you"]Whichever setting you choose, you can edit or remove any to-do on your list.
Related pages
- How Companion learns: what Companion learns from your timeline, and how.
- Personal AI roadmap: the learning features that are not yet built or only planned.
- Personas: the fictional people on this page, and the status labels.
- Scenarios: step-by-step walkthroughs of sharing and planning.
- Use cases and workflows: photo requests, calendar checks and agent delegation in detail.
- AI training, inference and jobs: the technical deep dive.
- Install and use Companion Memory: setup, every screen, and the API.