Every AI assistant reads your email. The question is how.
An AI assistant that does not read your email cannot help with your day. What matters is not whether it reads, but what it stores and for how long.
There is a question that comes up in every conversation about personal AI assistants: "does it read my email?". The honest answer is always yes. An assistant that does not read your email cannot tell you that you have a deadline on Friday or that your client has been waiting two days for a reply.
The right question is not "does it read it?". It is "what does it store?".
Reading is not storing
A well-built assistant reads the email once, extracts what it needs (who wrote to you, what they ask, whether there is a date) and discards the literal content. It does not store your email on its servers. It does not train a model on your messages. It does not need your Gmail password: it uses OAuth, a revocable permission that gives away no keys.
This is exactly what a human secretary does: read the letter, write down "call García on Thursday because he is asking for a quote", and return the letter to the mailbox. No copy kept, nothing taken home.
What it does need to remember
For the assistant to improve over time, it needs to keep a minimal profile: which senders are frequent, which types of emails you tend to ignore, and which always interest you. That is not the email: it is a statistical summary of your habits. It is the difference between storing all your emails and storing a note that says "almost never reads fintech newsletters".
That profile should be visible, editable and deletable. If you cannot see what the assistant has learned about you, you cannot trust that it has not learned too much.
How long things are kept
Retention is the other topic nobody explains. An assistant that says "we don't store your data" but keeps access logs for three years is cheating. An honest assistant says: "the daily summary is deleted after 48 hours; the habit profile is deleted if you revoke access; debug logs are purged every 30 days".
It is not complicated. It is transparency.
The model does not see your data
The last important nuance: the language model that generates the summary does not need to — and should not — access your data directly. A good design works like this:
1. The server reads your email with your OAuth permission.
2. The server extracts the signals (sender, subject, date, request).
3. The server sends the signals to the model as context, without the original text.
4. The model generates the summary without ever seeing the body of the email.
That is not an implementation detail. It is an architecture decision that protects the user without losing quality.
At Madrugo, that is exactly what we do: the model reads a normalised JSON, not your email. You can read more about how we handle memory at what an assistant that reads your email knows about you.