> ## Documentation Index
> Fetch the complete documentation index at: https://docs.inboxmate.psquared.dev/llms.txt
> Use this file to discover all available pages before exploring further.

# Learning from your edits

> How InboxMate turns your draft edits and declines into reusable guidance for future AI replies, and how to review or remove what it learned.

Every time you clean up an AI draft before sending it — or decline a suggestion and say why — InboxMate treats that as a lesson. It's not just logged for you to browse later: the correction is distilled into a short, reusable instruction and quietly fed into future drafts for similar emails, so the same mistake doesn't keep coming back.

This is different from [categorizer corrections](/user-guide/inbox/decisions), which retrain how emails get *sorted* and fade out after 30 days. This learning loop retrains how replies get *written*, and it never expires — a lesson sticks until you remove it yourself.

## How a lesson is created

Two moments in [Entscheidungen](/user-guide/inbox/decisions) trigger this:

<CardGroup cols={2}>
  <Card title="Approve with edits" icon="pen-line">
    You change an AI-drafted reply before sending it. InboxMate compares what was suggested against what you actually sent.
  </Card>

  <Card title="Decline with a reason" icon="circle-x">
    You reject a suggestion and give a reason (e.g. "zu förmlich, wir duzen diese Kunden"). The reason itself is the signal.
  </Card>
</CardGroup>

Small, low-signal edits — fixing a typo, tweaking a single word — are filtered out automatically; they're not worth turning into a rule. Anything more substantial goes to an AI classifier that sorts the change into one of five kinds and writes a short, reusable guidance instruction from it:

| Kind (German label) | What it covers                                                                |
| ------------------- | ----------------------------------------------------------------------------- |
| **Ton**             | Salutation, register, politeness (du/Sie, formal/casual, greeting or closing) |
| **Fakten**          | Wrong or missing facts — prices, hours, names, links, deadlines               |
| **Struktur**        | Length, ordering, paragraphing, what was added or cut                         |
| **Signatur**        | Sign-off / signature                                                          |
| **Sonstiges**       | Everything else                                                               |

The raw before/after text and the internal instruction are never shown back to you — only a short, plain-language German summary of what was learned is kept for the UI. If you correct the same kind of thing again, InboxMate recognizes it's the same lesson and just bumps a counter instead of storing a duplicate.

## Where lessons apply

Each lesson is scoped to your account, the **Postfach** (mailbox) it was learned in, the category the email was in, and — when relevant — the sender's domain. A tone correction learned generally will inform drafts broadly; a fact correction tied to `@lieferant-x.at` will mainly resurface for mail from that domain. When InboxMate drafts a new reply, it looks up the lessons most relevant to that email and quietly applies them — you won't see them called out in the draft, but the wording should reflect what you taught it.

## Reviewing and removing what it learned

Open **KI trainieren** in the sidebar, then the **Gelernt** tab (`/knowledge?tab=learned`). Each entry shows:

* A plain-language summary of the lesson
* A colour-coded kind chip (Ton, Fakten, Struktur, Signatur, Sonstiges)
* The sender domain, if the lesson is scoped to one
* A **{count}×** badge once the same lesson has been confirmed more than once
* How long ago it was learned

<Info>
  Nothing here shows the original email text or the exact wording InboxMate learned to say instead — just the customer-facing summary. This keeps the list readable and avoids surfacing sender-specific phrasing out of context.
</Info>

Click the trash icon on any entry to remove it. This is a soft delete: the lesson stops being applied to drafts immediately, and the action can't be undone from the UI.

<Tip>
  If a draft starts leaning on an outdated instruction (a price that changed, a tone you no longer want), find it in **Gelernt** and delete it — the next draft for that scope goes back to relying on your knowledge base and agent personality alone.
</Tip>

## Where this fits

Both this learning loop and inbox categorization corrections feed into how self-sufficient InboxMate becomes over time — tracked on [KI-Reife](/user-guide/inbox/ai-maturity). Draft edits specifically improve *what* the AI writes; categorizer corrections improve *where* an email ends up and *which* action fires.