> ## Documentation Index
> Fetch the complete documentation index at: https://agents.concierca.ch/llms.txt
> Use this file to discover all available pages before exploring further.

# How Your Workspace Learns

> The nightly curator: how repeated successes and failures become proposed skills and knowledge — always with your approval.

Your workspace can learn from its own work. When self-learning is enabled, a **curator** — a special background agent — watches what your agents do and turns recurring patterns into proposals: reusable skills, rules, and knowledge entries. It proposes; **you decide**.

## What the curator does

Think of it as a colleague who looks over your agents' shoulders and takes notes:

* *"We run this workflow successfully again and again — let's write a playbook."* → a **skill proposal**
* *"We keep failing on the same kind of error — let's write a rule against it."* → a **rule proposal**
* *"This note is durably important — let's keep it as knowledge."* → a **knowledge proposal**

It builds nothing on its own and decides nothing on its own. It observes, condenses, and proposes.

## What it learns from

The curator reads the **log of the work**, not your business data:

| It reads                                                  | It does not read                      |
| --------------------------------------------------------- | ------------------------------------- |
| Agent runs — what ran, and whether it succeeded or failed | The business data those runs touched  |
| The decision log — decisions agents recorded              | Anything outside your company         |
| Memories — the agent memory layer                         | Anything outside its department scope |

It sees *that* an agent summarized a report; it never re-reads the report's contents. And everything is scope-filtered — the curator only ever sees its own company and department.

## How a pattern becomes a skill

1. **The nightly sweep** groups recent runs: failures by error type, successes by workflow.
2. **The rule of three:** a pattern must occur at least **three times** before it is proposed. One-offs are ignored — the curator is deliberately conservative.
3. **Deduplication:** memories are de-duplicated and marked as processed along the way — archived, never deleted.
4. **A proposal is written** — pending, inactive, and invisible to agents until reviewed.
5. **You review it** in **Skills Review** (Sidebar › Knowledge › Skills Review, admins): approve to turn it into a real skill or knowledge entry, or reject it. Only an approved proposal ever becomes active.

<Note>
  A proposal is **not** a skill. Until a human approves it, nothing the curator writes is usable by any agent. That is the governance core of the whole feature: the machine suggests, the human decides.
</Note>

## Ownership

Approved skills and rules belong to a **scope**, not to a person: department-scoped proposals become usable by every agent in that department; global ones by everyone. Personal facts the curator captures go into user-scoped memory and stay private to that user.

## Turning it on

One switch: **Settings › Preferences › Self-Learning (Company Brain)** *(Company Admin)*. Enabling it provisions the nightly learning sweep plus a knowledge summarizer per active department; the setting shows an **estimated weekly model budget** before you commit. Disabling turns the learning agents off without deleting anything they produced.