Personal AI on your own hardware
The idea
Companion is being built to think in two ways. Quick checks will settle everyday questions in a moment, and a full language model will take the ones that need real thought. The language model already runs on your box through Companion Hub, and a first quick check is ready for apps and agents to use. The small models that will make those quick checks your own, trained on your own examples, are not built yet. Your records stay on hardware you own, and the models that learn from them run there too.
Two ways Companion thinks
This diagram shows how a question is meant to move from the quick check to the language model, and which parts exist today.
flowchart TB
Q["A question from an app or agent<br/>such as: is this a commitment?"] --> C["Quick check by meaning<br/>works today"]
C -->|"sure"| A["Answer"]
C -->|"unsure: hand-off not yet connected"| L["Full language model on your box<br/>works today"]
L --> A
P["Small models trained on<br/>your own examples<br/>not yet built"] -.->|"will make the check personal"| CA quick check compares the meaning of a question with each possible answer, using the embedding model on your box. It suits questions with a short list of answers: yes or no, pick one, give a score, or put a list in order. When the check is sure, its answer stands. When it is unsure, the question is meant to go to the full language model, which takes longer and can reason its way to an answer.
- Works today: the full language model, through the local models Companion Hub runs on your box. Chat and other open questions go to it directly, and Companion also uses it to sort your activities and to find to-dos.
- Works today: the quick check, for apps and agents you connect to your box. Nothing in Companion calls it on its own yet. You can switch it off under Settings → Model.
- Not yet: the hand-off from an unsure quick check to the language model. For now, a simple word-matching check fills that step.
- Not yet: small models trained on your own examples, which will make the quick checks personal.
What ships today
Companion already learns from your own history, and that learning happens on your box.
A language model on your box sorts your days into broad activities: work, exercise, social time, errands, leisure, travel and sleep. From those, Companion learns when you usually do what. The Calendar page, under Planning, shows the result behind the hours of each day as a predicted rhythm for the week ahead.
Each night, your box also updates a small personal add-on for an older, larger model of your routine, using the day before. Since Companion Memory 2026.11.11, each update is saved as a new version, a failed run is never saved, and your box keeps the three versions before the current one. Nothing in the product rolls back to them yet. The forecast you usually see comes from simpler models, explained below, so this nightly update doesn’t change it.
- Kinds of activity
- 7
- Work, exercise, social, errands, leisure, travel and sleep. Anything it can't place is marked unknown.
- from vocab_registry.py
- Personal update
- Nightly
- Your box updates the add-on for its older routine model from the day before.
- from nightly_finetune.py
- Earlier versions kept
- 3
- Each nightly update keeps the three before it, from Companion Memory 2026.11.11.
- from train_io.py
Companion also looks through your messages for things you need to do. What happens next depends on the trust level you chose at setup, which you can change under Settings → Autonomy. On Balanced, the default, the to-dos it is confident about join your task list on their own, and the rest wait under Proposed for you to accept or dismiss. Putting a to-do on your calendar still asks you first. On Cautious, every to-do waits under Proposed. On Hands-off, Companion adds and schedules to-dos on its own.
Capability status
This table lists each part of Companion’s personal AI and where it stands.
| Capability | What it means for you | Status |
|---|---|---|
| Your timeline on your box | Messages, calendars, photos, documents and location history that you import or capture are stored on hardware you own. None of it is sent to a cloud service to be learned from. | Works today |
| Activities sorted for you | A language model on your box sorts your days into seven broad activities, such as work, exercise and sleep, or marks it unknown. | Works today |
| Predicted rhythm | The Calendar page, under Planning, shows how your coming week is likely to go, learned from your own history. | Works today |
| Nightly personal update | Each night your box updates a small personal add-on for an older routine model, from the day before. Since Companion Memory 2026.11.11 it keeps the three earlier versions on disk. Nothing in the product restores one yet. | Works today |
| A check on the older model | Companion compares the older model’s past forecasts with patterns from your usual week, and sets the model aside when it doesn’t clearly beat them. | Works today |
| Suggested to-dos | Companion finds to-dos in your messages. On Balanced, the default, confident ones join your task list on their own and the rest wait in Proposed. On Cautious, every one waits in Proposed. | Works today |
| Quick check by meaning | Apps and agents you connect can ask yes-or-no, pick-one, score and ranking questions. Nothing in Companion calls it on its own yet. | Works today |
| Forecast page | A page of its own for the forecast. It stays off unless the box’s owner turns it on in the server’s configuration. There is no switch for it in the app. | Behind a setting |
| Keyword shortcut for activities | A shortcut that sorts some activities by matching keywords instead of asking the language model. It is off by default from the release after Companion Memory 2026.11.11 (that release has it on), and the box’s owner can turn it on. | Behind a setting |
| Hand-off to the language model | When the quick check is unsure, it will pass the question to the language model. A simple word-matching check fills that step for now. | Not yet |
| Quick-check models trained on your examples | Small models, trained on your box from your own examples, that answer questions like “is this a commitment?” or “is this worth remembering?” quickly. The training code is a first draft, and nothing is trained yet. | Not yet |
| Training on another device you own | Your box will be able to hand training work to a computer you own with a bigger graphics card, through Companion Hub. Hub has the first parts of this, but it can’t yet run a training job from start to finish. | Not yet |
| Checks for every new model | Every newly trained model will have to beat simple guesses, and show confidence that matches how often it is right, before Companion uses it. | Not yet |
| Tone add-on | Companion will learn your writing style, so drafts sound like you. | Planned |
| Commitments staged for you | Companion will spot commitments in your conversations and stage a calendar event or task for you to confirm. | Planned |
| Insights at the right moment | Companion will judge whether an insight is worth interrupting you for. | Planned |
| Smarter search and a cleaner memory | Personal models will rank search results, spot duplicate contacts and filter out noise before it is stored. | Planned |
Roadmap
The roadmap groups this work into three stages, without dates.
flowchart TB
N["Now<br/>Routine forecasts<br/>Nightly personal learning<br/>Suggested to-dos"] --> X["Next<br/>Quick-check models trained on your box<br/>Training on another device you own"]
X --> L["Later<br/>A tone add-on<br/>Commitments staged for you to confirm<br/>Insights that interrupt only when worth it<br/>Smarter search ranking"]- Now: routine forecasts on the Calendar page, nightly personal learning that keeps earlier versions on disk, and suggested to-dos.
- Next: quick-check models trained on your own box from your own examples, and training on another computer you own, such as a desktop with a bigger graphics card, through Companion Hub.
- Later: a tone add-on so drafts sound like you, commitments staged for you to confirm, insights that interrupt you only when it’s worth it, and smarter search ranking.
Why on your own hardware
Your memory holds your messages, photos and the places you have been, so Companion keeps it, and the models that learn from it, on a box you control. Here is how a default setup compares with a typical cloud assistant.
| A typical cloud assistant | Companion | |
|---|---|---|
| Where your data lives | In the company's cloud | On your box, at home or at work |
| Where the models run | On the company's servers | On your box, through Companion Hub, unless you choose a cloud model |
| Who trains models on your data | The company, as its terms allow | Your box, for your own personal models |
| Who can read it | The company's systems, under its policies | You, and the people you share with, under rules you set |
| When the internet is down | It needs a connection to answer | The models on your box keep working. Lookups that reach outside, such as map tiles, need the connection |
A few background jobs do reach outside your box, such as crash reports, which you can turn off, and place lookups. What leaves your appliance lists each one and whether you can turn it off. If you choose a cloud model in Companion Hub, the requests you send to it go to that provider. AI and inference describes the local engines Hub can run instead.
How we’ll know the models are good
Before Companion relies on a model it has trained, the model should prove itself on your own history. Two questions decide it. Does the model beat a simple guess? And when it says it is sure, is it usually right?
- Not yet: both checks for every newly trained model, including the quick-check models. A model that fails either check will not be used.
Part of this runs today, for the forecast on the Calendar page:
- Works today: while your history is short, the forecast follows simple patterns from your usual week: what you usually do at that time on that day of the week.
- Works today: a check on the older routine model that your box updates each night. Companion scores its past forecasts against what actually happened. When they don’t clearly beat your usual-week patterns, or its confidence is clearly less reliable than theirs, Companion sets the model aside.
- Not yet: the same check for the small model behind most forecasts once there is enough history. Companion refits that model from your history and uses it without this check, so the forecast you see has not always beaten a simple guess.
Related pages
- How Companion learns: how Companion learns from your own history.
- At work and at home: Companion’s personal AI in everyday use.
- AI training, inference & jobs: the technical deep dive.
- System architecture: how Companion Memory is put together.
- Install and use Companion Memory: setup, every screen, and what leaves your appliance.
- AI and inference: the local models Companion Hub runs.