1 · Your agent's home
TL;DR. A server in your name. Not shared with anyone else.
Each agent runs on a dedicated server we set up in your name. Other users' agents never share that server. Your agent's memory is encrypted at rest.
2 · What we store
TL;DR. Conversations live on your server, not ours.
- On your server: your conversations, your agent's long-term memory, the routines it runs, and a short technical log of what it did.
- In our control plane: your AgentHut account email, subscription state, server status, the basics needed to operate the platform.
- Not stored in our shared control-plane database: the contents of your chats and memories. Connector content can pass through the control plane when needed to deliver a feature to your dedicated server or model provider; the Google-specific processing and storage are described below.
2a · Your memory is inspectable — not a black box
TL;DR. See what your agent knows, correct it, or make it forget — because the memory is yours.
- See it: your dashboard shows a live meter of exactly how much context your agent holds — per source, counted, timestamped. Or just ask it: "what do you know about me?"
- Correct it: tell your agent a remembered fact is wrong and it updates the record.
- Forget it: tell it to forget a topic, a person, or a whole source. The deletion happens on your server, where the data lives.
This works because of architecture, not policy: the memory sits in a database on your dedicated server. We can't read it — so showing it to you, and only you, is the natural default.
3 · Model providers
TL;DR. The model API only sees what it needs to answer.
Your agent uses model APIs (Anthropic, OpenAI, Mistral, Gemini, or your own keys). API requests leave your server only when needed to answer a prompt, and providers' data-retention policies apply to that traffic. With your own keys, your zero-retention agreement carries over.
4 · Permissions
TL;DR. Anything that touches the world asks first.
Sensitive actions — sending messages, paying, scheduling, deleting — stage themselves and ask before they run. You can revoke any permission, any tool, any time.
4a · Google Workspace data
TL;DR. Gmail is accessed only when you ask Ermes to use it. It is never sold, used for ads, or used to train general-purpose AI models.
In this Gmail-only production release, connecting a Google account lets Ermes access the Google account address and the specific Gmail data covered by the permissions shown on Google's consent screen:
- Gmail read/search: message metadata and content needed to answer the Gmail question you ask in the current turn.
- Gmail drafts: recipient, subject, thread, and body data needed to create a draft only after you approve its exact text. Ermes does not send email.
Relevant Gmail content may be transmitted to your dedicated agent server and to the model provider configured for your agent solely to perform that request. Connecting Gmail does not start a background import, scheduled mailbox scan, or automatic durable-memory save in this release. Ermes saves Google-derived content into durable memory only if you separately ask it to save that specific information. The connector's central database stores the connected account address, granted permissions, connection-health metadata, and an encrypted Google refresh token; short-lived access tokens and Gmail message content are not retained there.
We do not sell Google user data, use it for advertising, use it to determine creditworthiness, or allow humans to read it except with your affirmative permission for support, when necessary for security, or when required by law. We do not transfer it to third parties except to provide the features you requested, with your consent, or for the limited security, legal, or corporate-transfer cases allowed by Google's policy.
You can disconnect Google from the Connectors page. Disconnecting revokes the Google authorization where Google permits it and removes the stored connector credentials, which stops future access. Information you separately asked Ermes to save remains in your agent memory until you delete it, ask Ermes to forget it, or delete your account. Account deletion removes the dedicated server within 7 days. You can also revoke Ermes directly from your Google Account connections.
Ermes's use and transfer of information received from Google APIs adheres to the Google API Services User Data Policy, including its Limited Use requirements.
5 · Export & delete
TL;DR. One-click delete. JSON export anytime.
You can export the full memory + routine library as JSON. You can delete everything in one click; we wipe the server within 7 days.
6 · Who at AgentHut can see what
TL;DR. By default, nobody. Read-only on request, auto-expires.
By default, AgentHut staff have no access to your server. When you ask for support, you can grant temporary, read-only access; we close that access automatically after 24 hours.
7 · Sub-processors
TL;DR. Short list. 30-day notice before adding any.
We list the third parties we rely on (server provider, DNS, payment, email) on our sub-processors page. We'll email you 30 days before adding any new sub-processor.
8 · Legal & contact
TL;DR. US controller (Delaware). GDPR rights apply.
Data controller: AgentHut Labs Inc. (Delaware, USA). For privacy questions, use the form on our contact page or write to [email protected]. Your GDPR rights (access, correction, deletion, portability) apply to data we process about you, regardless of where the company is incorporated.
9 · Cookies
TL;DR. Strictly necessary only. No third-party trackers.
The marketing site uses a couple of strictly-necessary cookies to remember your session and your preferences. We don't run third-party advertising or analytics trackers. If we ever add analytics, we'll ask first and let you say no.