# Learning

How a goal learns from its own work, what it changes on its own, what waits for you, and how to undo it.

Canonical page: https://app.auto-lab.ai/docs/brain/learning

Your goal gets better at its work by looking back at it. After corrections, long tasks and quiet spells, and again
each night, the agent reviews what happened and saves what it learned. You see all of it in **Brain › Learning**.

Small changes that are easy to reverse, such as a note or a lesson, take effect at once and each has **Undo**.
Changes to what a person wrote, or to the rules the agent follows, could change its behaviour in ways you did not
ask for, so a person decides on those.

This page describes goals that use Chat. For older goals that run in sessions, see
[Older session goals](#older-session-goals).

## When the goal reviews its work

A review runs in the background after the agent has replied. It reads the conversation since the last review.

| When | What starts a review |
|---|---|
| You correct the agent | Right after that turn |
| A long task finishes | A turn that used many tools |
| A long conversation | Every 10 of your messages |
| Something went wrong | A turn failed, stalled, ran out of steps or hit something in its way |
| The agent worked on its own | A turn that ran with no person involved |
| A conversation goes quiet | Checked hourly. The review runs once nobody has written for two hours |
| Every night | Around 3:30 in the goal's time zone, if the day had any activity |

Besides corrections and the nightly pass, a goal runs at most six reviews a day. Each review saves at most three
things and proposes at most three. If you write while a review is running, it stops without saving anything and
runs again after your turn.

## Choose how much it learns

Set the learning level in **Settings › Autonomy**, under **Learning**.

| Level | What happens |
|---|---|
| **Off** | Nothing is learned. Skills and memory still work. |
| **Ask me first** | Every change waits for your approval in **Needs your decision**. |
| **Learn on its own** | The default. Notes, preferences you stated, lessons and its own skills are saved at once, each with **Undo**. It asks before changing anything a person wrote. |

The top of **Brain › Learning** shows the current level, for example "Learning on its own", with a link to
**Autonomy settings**. Changing the level needs the goal's admin role.

## What it changes, and what it never changes

At **Learn on its own**, these are saved at once, each with **Undo**:

- A note about what happened, added to the goal's timeline, people, conversations or workstreams. It only adds;
  it never rewrites a note.
- A preference you stated in the conversation, saved with your own words.
- A lesson from something that went wrong, added to the playbook that every run reads.
- A skill of its own, named `learned-…`, or an edit to one. First Auto Lab replays recent turns with the new
  skill. If anything gets worse, or the replay cannot run, the skill waits for you instead.

These are always proposed as a change request, at every level:

- Replacing or removing something already in memory
- Knowledge the agent worked out, rather than heard from you
- The objective
- Edits to a skill a person wrote
- One small edit to the [goal instructions](/docs/brain#goal-instructions): one line added, reworded or removed

Learning never changes the agent's own instructions (its prompt), its tools or the goal manifest (`kortix.yaml`).

Anything you undo or decline is remembered, and the agent does not save or propose it again.

## See what it learned

In Chat, a line appears after a review, for example "Learned: Send invoices on Fridays" with **Undo**, "Learned 3
things" with **See all**, or "1 suggestion waiting for you" with **Review**.

**Brain › Learning** shows these sections, top to bottom:

| Section | What it shows |
|---|---|
| **Needs your decision** | Open change requests from learning. Each row names the proposed changes and, for a skill, the replay result, such as "Replay: no regressions". |
| **Learned on its own** | What the agent saved without asking, newest first: why (its reason, or your own words), when, and what started the review. **Open** shows the file in the Brain. |
| **Skills** | Every skill, with how often the agent opened it and when last. Learned skills come first. |
| **Is it helping?** | Corrections per week, which should fall as it learns, and how much was undone or declined. |
| **What changed** | Every edit to agents, skills, memory and tools that took effect, by day, with who made it. |
| **Outcome trend** | The number the goal tracks over time, with accepted changes marked on it. |
| **Reviews** | Each time the agent looked back at its work, and the result. |

Further down you find **What got in the way** (the agent's own notes on what was missing or wrong), **Replay
results**, **Lessons** and **Learned this week**. Anyone who can open the goal sees this tab. Some sections, such
as **What changed** and **Lessons**, need the goal's admin role.

## Decide on a proposal

### Open it
In **Needs your decision**, select **Review** on the row.

### Read the change
Check what would change. For a skill, check the replay result.

### Decide
Select **Apply changes** to accept it, or **Dismiss** to decline it.

The same change requests also appear in **Needs you** on the Overview and in the Review Center. See [Approvals and autonomy](/docs/goals/approvals).

## Undo something it learned

Select **Undo** in either place:

- In Chat, on the "Learned: …" line.
- In **Brain › Learning**, on the row under **Learned on its own**.

Undo removes exactly what was added and records the removal in the goal's history. The row stays, marked
**Undone**. If the file changed since, Undo cannot run safely and says "This changed since; edit it in Brain." Open
the file and edit it yourself. Undo needs the goal's admin role.

## Look after its skills

The **Skills** section helps you keep the agent's skills useful:

- **Pin** a skill to keep it. Pinned skills are never retired.
- **Archive** a skill to take it out of the list the agent chooses from. **Restore** brings it back.
- A learned skill nobody opened for 14 days is marked **Stale**. After 30 days it is archived, and you can undo
  that like any other learned item. A skill a person wrote is only ever marked stale.

To edit a skill, see [Agents and skills](/docs/brain/agents-and-skills#edit-a-skill).

## Older session goals

Goals that run in sessions instead of Chat learn through a **Daily reflection**. Once a day it reviews recent
finished sessions and proposes changes as change requests. Nothing applies on its own.

- It is off by default. Turn it on with the **Daily reflection** switch in **Settings › Autonomy**.
- It needs five completed sessions before you can turn it on.
- While it is off and the goal qualifies, the Overview's **Needs you** shows a **Turn on daily learning**
  suggestion.

On these goals, **Brain › Learning** shows the reflection's status at the top and lists each run under
**Reviews**. Open a run to read its findings and **Accept** or **Reject** each one.