A mailbox holds most of what a workspace needs to know about a person's work:
who they write to, what is still open with each of them, and how they write.
Reading all of it again every morning is how a job that should be routine
turns into a bill nobody expected.

Learn How I Talk reads the email that matters into memory, learns how the
person writes from what they send, and keeps a mark of what it has read. The
next run starts where the last one stopped.

## What it reads

Sent mail from the last three months, what was starred, what matters from the
last month, and the threads with people already in memory. Newsletters,
notifications and no-reply senders are skipped on sight. Each person and
company in a thread gets a page with what was discussed and what is open, and
the person's own writing is read again for voice: length, greeting, sign-off,
the phrases they reach for and the ones they never use.

## The run says what it will cost before it reads

Every run states, before it reads a message, how many messages it will touch,
from the count the mailbox returns with the first search. On Personal nothing
is charged to GBrain for a run, because the run spends the AI's own plan.
Saying **everything** covers the whole mailbox instead of the last three
months, and asks first.

## The second run reads only what is new

A workspace keeps a checkpoint for each connected account and each search: the
day that search reached. A run starts from that day, so it lists what arrived
since and skips every thread already read, however many times it runs. That is
why a second run costs a fraction of the first. A run can stop at any point and
carry on from the same place, and one conversation is read once rather than
again every morning.

## An AI that may write

The skill files pages, entries and links, so the AI running it needs
**Read & Write** on the memory, not Read. [Use the memory from other
AIs](/docs/workspace/memory-anywhere) has the connection steps, and [Connect an
AI](/docs/tools/assistants) has the setting per AI.

On Personal the workspace runs nothing on its own. The AI runs the skill as one
of its scheduled tasks, from a prompt:

```
Use my GBrain skill "Learn How I Talk" every morning at 7. Set this up as a
scheduled task that repeats.
```

Change the hour, or say every weekday, in the words. The workspace's page for
that AI, under **Integrations**, has the same prompt with a Copy button, and
so does the skill's page.

## What it leaves alone

It only reads: it never sends, drafts, archives, labels or deletes anything.
Health, therapy and family medical detail stay out of every page and every
dated entry. A stop is safe, because the checkpoint on disk is the progress,
not the conversation.

## Next

[Connect an AI](/docs/tools/assistants), give it Read & Write, and paste the
prompt above. The first run reports what it read and what it filed; ask it for
a draft reply to anyone it found.