How Companion learns
Companion Memory keeps your timeline on your box, the hardware you own. That timeline holds the messages, calendar entries, places, photos and documents you import or capture. From it, Companion learns the usual shape of your days: when you tend to work, exercise, run errands or see friends.
The learning happens on your box, and what Companion learns stays there. It uses what it learns to forecast your day and to help you plan. You decide how much it does on its own, and you can plan or dismiss anything it suggests. Some of this is still being built, and the table near the end gives each part a status.
The learning loop
The diagram shows how your days become forecasts and how you feed back in.
flowchart LR
A["Your apps and devices"] --> B["Your timeline, on your box"]
B --> C["Daily activities"]
C --> D["Models of your routine"]
D --> E["Forecasts"]
B --> F["Suggested to-dos"]
E --> G["You decide"]
F --> G
G -->|"what you do and finish"| BYour apps and devices
Your timeline fills from the archives you import, the accounts you connect, and capture apps such as the browser extension and the mobile apps.
Your timeline, on your box
Everything lands in one timeline on your box. Learning reads from this copy.
Daily activities
Once a day, Companion sorts new items into broad activities: work, exercise, social, errands, leisure, travel and sleep. Simple rules handle clear cases, such as a workout or a trip. Calendar entries and other text go to the language model on your box. If Noor, one of the fictional personas, has client meetings most weekday mornings, those mornings are likely to be sorted as work.
Models of your routine
Companion keeps small models of your routine on your box. A small model is fitted from your history once for each day it forecasts, and reused for that day. A larger personal model trains on your box and is updated once a day from the day before.
Forecasts
A forecast shows what you are likely to be doing at each time of day. The Calendar in Planning shows your usual week next to the events you have booked, and when you pick a time, its Predictions panel says what usually happens then. The Prediction card on the Widgets page shows the next forecast, and Chat can read your forecast when you ask about your day. A forecast never adds anything to your calendar or tasks by itself.
Suggested to-dos
Separately, Companion reads your recent messages and email for things you need to do. On the default Balanced setting, to-dos it is confident about are added to your tasks, and the rest wait in the Proposed column of Planning → Tasks. In Settings → Autonomy, Cautious keeps every one waiting for you, and Hands-off adds them and places them on your calendar.
You decide
For each suggestion, choose Plan, Schedule or Dismiss. Companion checks new suggestions against the ones you dismissed, so the same to-do doesn’t keep coming back. What you actually do, including the tasks you finish, joins your timeline, and Companion learns from that next.
What happens overnight
Once a day, at a fixed time, Companion runs a nightly update for each account that has finished setup. The diagram shows its steps in order.
sequenceDiagram
participant T as Your timeline
participant L as Language model on your box
participant M as Personal model
participant S as Nightly scores
T->>L: New items from the day
L-->>T: Sorted into activities
T->>M: The previous day's activities
M->>M: Save a new version, keep three older ones
T->>S: What actually happened
S->>S: Score past forecasts and the simple guesses
S-->>M: Used only while it beats the better guess- Companion sorts the day’s new items into activities. A big import is sorted a little each night, newest first.
- If you have no personal model yet, or it no longer fits your activities, Companion trains a new one from your history.
- It updates a small add-on to the personal model from the previous day and saves it as a new version. Since Companion Memory 2026.11.11 the three older versions are kept, and a failed run leaves the last good one in place.
- It records what you actually did in the slots it forecast, and scores past forecasts and the simple guesses against that.
- It rebuilds your usual weekly rhythm, which the Calendar shows.
Putting back an earlier version is not in the app yet, and nothing switches back on its own.
Which forecast you see
Each time Companion makes a forecast, it picks from three sources:
- Simple guesses from your usual week. What you usually do at that time on that day of the week, plus a version that also follows which activity usually comes next.
- The personal model. Companion considers it once it has about two weeks of your history. It uses the model only if its latest score beat the better simple guess by a clear margin, without being noticeably worse at judging how sure it is. Otherwise it shows that guess. A model with no score yet is used without this check.
- The small model. It weighs the time of day, the day of the week, what you were doing at that time yesterday and a week earlier, and what you have done lately, with recent days counting most.
Once you have a few days of history, the forecast you see usually comes from the small model. Companion uses it whenever there is enough history to build it, and it skips the check above. Before then, you see the simple guess from your usual week. So the nightly update changes the personal model, which today seldom decides the forecast you see. Quality checks that every newly trained model must pass are written but not connected to anything yet.
Where your data goes
The diagram shows where learning happens. The dotted line is not built yet.
flowchart LR
subgraph box["Your box"]
T["Your timeline"] --> J["Learning jobs"]
J --> M["Models of your routine"]
M --> F["Your forecast"]
end
subgraph net["Your private network"]
D["Another device you own"]
end
C["Cloud services"]
J -.->|"Not yet: bigger training jobs"| D
J --x|"no training here"| CYour models train, run and are stored on your box. Companion doesn’t send your timeline to a cloud service to train them.
Sorting items into activities relies on a language model. It uses the model your box is set up with, which in a standard setup runs on the box itself. If you have paired other boxes in Companion Hub, one of them may answer instead, over your private network. A cloud model is used only if you added a cloud provider in Companion Hub and no model on your own hardware is ready.
Not yet: training on another device you own, such as a desktop with a bigger graphics card. Companion Hub is being built to send bigger training jobs to that device over your private network, so the work stays on hardware you own. Today it can’t run a training job there from start to finish.
What works today and what’s coming
The four status labels mean the same as on the personas page.
| What | Status |
|---|---|
| Your timeline and your models stay on your box, and training never runs in a cloud service | Works today |
| Sorting each day into broad activities, with simple rules and the language model on your box | Works today |
| A nightly update that learns from the previous day, saved as a new version with the three before it kept (since Companion Memory 2026.11.11) | Works today |
| Using the personal model only while it beats a simple guess from your usual week | Works today |
| A small model, fitted once for each forecast day, that supplies most forecasts and skips that check | Works today |
| Your forecast on the Calendar, on the Widgets page and in Chat | Works today |
| Suggested to-dos from your messages and email, following your Autonomy setting | Works today |
| A quick-decision service that apps and agents on your box can call | Works today |
| The Forecast page, a full view of your forecast, which needs a setting on the box | Behind a setting |
| A switch that pauses learning | Not yet |
| Handing the quick-decision service’s unclear cases to the language model | Not yet |
| Putting back an earlier version of the nightly update from the app | Not yet |
| Small decision models trained on your own examples, for questions like “is this a commitment?” | Not yet |
| Training on another device you own through Companion Hub | Not yet |
| Learning your writing style so drafts sound like you | Planned |
| Spotting plans in your conversations and staging calendar events for you to confirm | Planned |
| Insights that interrupt you only when they are worth it | Planned |
| Using decision models to rank search results and to filter noise before it is stored | Planned |
| Quality checks that every newly trained model must pass before it is used | Not yet |
The quick-decision service answers yes-or-no and pick-one questions by comparing meaning. When the answer is unclear, it falls back to a simple check of shared words. It doesn’t learn from you, and no built-in feature calls it yet. Its switches, on by default, are in Settings → Model under Memory governance & decision engines.
Questions people ask
Can I turn learning off?
Not yet. There is no switch that pauses it, and the nightly update runs for every account that has finished setup. You can choose how much Companion does on its own in Settings → Autonomy. On Cautious, every to-do it finds waits for you.
What if it gets my routine wrong?
A forecast is only a guess, and it never changes your calendar or tasks by itself. Recent days count for more than older ones, so the forecast follows a new routine as it repeats. If an event didn’t happen, delete it. Deleted events, and events your calendar marks as cancelled, drop out of your activity timeline at the next nightly update.
I imported years of history. Why is my forecast thin?
Sorting history into activities takes time. Companion works on it for a set time each night, starting with your most recent months, and older history follows over later nights.
Does it need a GPU?
No. In the standard setup, your models train and run on the box’s processor. The biggest job is training the personal model, on the first run and again when your activities change a lot. Each nightly update learns only from the day before.
Related pages
- Personas: the fictional people in examples.
- Scenarios: walkthroughs of what works and what is missing.
- At work and at home: Companion Memory at work and in a household.
- Personal AI roadmap: what is being built next.
- AI training, inference and jobs: the technical deep dive.
- Install and use Companion Memory: setup, screens and settings.