An AI tool can get better at your business without any model ever being trained on your data. It does that through three things that live outside the model: the documents you connect, the memory it builds from your feedback, and the skills you approve. All three are visible and can be undone. Training is the opposite: your data is folded into model weights you cannot see, and cannot take back.
What «training on your data» actually means
Training means a provider uses your prompts and documents, alongside millions of others, to adjust the model’s parameters. Once that has happened, there is no deleting your contribution. Some consumer AI services do this by default. Business terms differ by vendor and plan. The first thing to check with any vendor is the sentence on training, in writing, for your account.
Nrth OS does not train models, and no model is trained on your data. That is the easy part.
How it learns anyway
| Mechanism | What it is | Where it lives | What you can do with it |
|---|---|---|---|
| Knowledge | The folders, libraries and files you connect | Synced from your Google Drive, SharePoint or uploads, indexed in your home region | Add, remove, see when each source was last synced |
| Memory | Short lines the assistant writes from your feedback («Use a table; put delivery terms last») | Two layers: company memory, shared, and assistant memory | Read, edit, delete each line; see what changed since last time |
| Skills | A task that worked, saved as a reusable recipe | The Skills area of your workspace | Approve, test, delete; see when each was last tested |
None of these change a model. The model receives, for each task, the instructions, the relevant knowledge, the memory lines and the skill, and does the work. The next task starts from the same text.
Why this is better than training, even if training were allowed
- It is inspectable. When a result is wrong because of something the assistant «learned», you can find the line and fix it. With training you can only retrain.
- It is reversible. Delete a memory line and it no longer goes into the next task. Delete a skill and the shortcut is gone.
- It survives a model change. Knowledge, memory and skills are text. If you switch models, or move to your own model endpoint, they come with you. Nrth OS is built model-agnostic.
- It is auditable. «Two memory lines changed since last time. Review» is a sentence a compliance officer can act on.
What feedback does
Every result has Good and Bad, and Bad lets you pick what was wrong, for example wrong content or wrong format, and write a short line about it. «Run again» takes your feedback with it, and feedback can become a memory line you can read and change. Feedback nobody gives teaches nobody anything.
What we don’t do
We don’t build a profile of your company that you can’t see. We don’t share memory between customers. More on the data page.
Next step
See how your data is handled, or apply for the pilot program and watch the first memory line get written.