EmAGI
Docs · index

One agent, several chats: what carries over

Diagram: one agent memory it answers from across different chats.

Your agent is one person, not a copy per conversation. That is the point of hiring an employee — and it is also the thing to set boundaries on, deliberately.

Is my memory shared between chats?

Partly, and the distinction matters. Every chat holds the history of that conversation — what was said there, in what order, what is still open. Above it sits one shared layer: durable knowledge about your business, your decisions and your standards. The agent can look across its own chats, and that is deliberate — it is why it stops asking the same questions after the first week, and why what it learned in the sales chat is there when you write about an invoice.

This is not isolation, and it should not be. Forbid an agent from looking into its own other conversations and you no longer have one employee — you have an agent without memory, meeting you again every Monday.

Then how do I keep something to one conversation?

Tell it, in that chat, in plain words: "what we discuss here stays here — never bring it up elsewhere." That is not a polite request that fades in a week; it becomes a standing rule for that chat and it is loaded every time the agent works there. The same works for style ("in this chat answer in English only"), for tone, and for anything you want the agent to do differently in one place.

If you ever see it carry something across that you wanted kept, say so once and ask it to write the rule down. An agent that keeps repeating the same slip has not written it down — insist that it does.

What if something must never mix at all?

Then the clean boundary is a separate agent for that work. Two agents share nothing — different worker, different memory, its own isolated server — so nothing can cross between them by accident. Per-chat rules are the flexible answer; a separate agent is the absolute one.

Can another company's agent see my data?

No. A different agent is a different worker on its own isolated server, with its own memory. Sharing memory between two of your agents is possible, but it is a deliberate action you take — never something that happens by itself. Details: security and your data.

Does it read every message in a busy group?

Only when it should. In a group you can set it to wake only when it is addressed — by reply, by mention, or by name — and to sleep through everything else. It is the right setting for any chat where people mostly talk to each other. See group chats and topics.

I sent five documents at once. What is the best way?

Say in one line what they are together: "here are five invoices for July, process them as one set and give me a single summary." An agent takes a task at a time, and one sentence turns five arrivals into one task with one answer instead of five separate passes. The same applies to a long list of questions — number them and ask for one reply.

Where do I see what it is working on?

Ask it: "what is open on you right now?" A healthy agent answers with the real list, per chat, and can tell you what it is doing at this moment even mid-task.