When should an AI agent start a new chat?
Talk to an AI agent long enough and the conversation turns into a junk drawer: last week's flight search, a half-finished invoice, a joke you made on Monday. Starting over feels like losing everything it learned. It doesn't have to. Here's how we split the two in nox.cool, and the small card that tells you what an agent is up to.

Key points
- A chat is the conversation; memory is what the agent knows about you. They don't have to be the same thing.
- On nox.cool, each agent keeps one ongoing chat, and a New chat button starts a clean one without wiping its memory.
- The old chat is kept, named by its first message, and you can go back and carry on in it.
- Jobs still running move to the new chat, so their results still reach you.
- A card beside the chat shows the agent's status, its computer, its earlier chats, its routines and what it made.
- Start fresh when you change subject, when replies drift, or when an old instruction keeps getting in the way.
A chat and a memory are two different things, and an AI agent works better when you can reset one without losing the other. On nox.cool, every agent keeps one ongoing chat with you. Since today, a New chat button starts a clean one, and the agent still remembers everything it learned about you. The chat before it is kept, one click away.
What's the difference between a chat and an agent's memory?
A chat is the conversation: every message you sent, every reply, every step the agent took. The AI reads it again each time it answers, which is how it follows "now send that to Sam" without you repeating who Sam is.
Memory is what the agent has decided is worth keeping about you: that you run a small T-shirt business, that Dana is your printer, that you'd rather get one short email than three long ones. On nox.cool these are notes you can read, change or delete on the Memory page.
Most chat apps blur the two. Start a new conversation and the assistant knows nothing. Keep one forever and it knows everything, including things you'd rather it forgot.
A clean chat shouldn't mean a blank agent.
Why do long chats go wrong?
Three reasons, all ordinary.
They get slow and costly. The model reads the whole conversation before each reply. A chat that has run for weeks is a long read every single time. nox.cool already condenses long work into a summary, but a summary is still more to carry than a clean page.
Old instructions linger. "Keep replies short today" was meant for today. Three weeks later, a model reading the whole chat may still be keeping things short.
Topics bleed. You asked about flights to Lisbon on Monday and invoices on Thursday. On Friday, "find the cheapest one" could mean either.
People already sense this. They open a new window to "start fresh" with a chatbot, and lose everything it knew.
How does New chat work on nox.cool?
At the top right of any agent's chat there's a pen button. Press it and:
- The agent's greeting and an empty message box come up.
- The chat you had is put away, not deleted. It's listed under Earlier chats on the agent's card, named after the first thing you said in it.
- Any job still running moves to the new chat. When it finishes, its result reaches the agent with your next message there.
- Memory stays exactly as it was. Nox still knows your printer's name.

If the chat is already empty, the button does nothing: there's nothing to put away, and you won't end up with a pile of blank chats.
Can I go back?
Yes. Click an earlier chat on the card and it opens as you left it, marked Earlier chat at the top. Keep writing in it if you like; the agent picks up where that conversation was. It's still that agent, on the same computer, with the same memory.
When should you start a new chat?
There's no rule, but a few signs are worth acting on:
- You're changing subject completely. From planning a trip to chasing invoices.
- Replies keep mentioning something you've dropped. That's the old part of the chat leaking in.
- An earlier instruction is in the way. "Only use the work calendar" from a week ago, when you now want both.
- It's been running for weeks. A clean start is quicker for the agent to read and cheaper for you.
And when not to: if the next thing builds on what you just did ("good, now send it to the team"), stay. That's what the ongoing chat is for.
| Stay in the chat | Start a new chat | |
|---|---|---|
| The next step uses what you just did | ✓ | |
| A new, unrelated job | ✓ | |
| Old details keep creeping in | ✓ | |
| You want the agent to remember you | ✓ | ✓ (memory carries over) |
What's the card beside the chat?
Once there are several agents, each with its own computer, it helps to see one at a glance without opening anything. So every agent now has a card beside its chat, painted in its own colour:

- Its face and status. "Working now" with a green dot, or when it last did something.
- Call. Talk it through out loud, the same live call as anywhere else in nox.cool.
- Telegram or Slack. If you've linked one, it opens your chat with Nox there. If not, it takes you to where you connect one.
- Its computer. Ready, working, or asleep (it wakes when needed). A click shows its screen.
- Earlier chats. The ones New chat put away, newest first.
- On its own. The routines it runs by itself, with when: "Every weekday at 08:00".
- Made. The newest files it made for you.
Each card shows only that agent's things. Open your researcher and you see its routines and its reports, not Nox's.
Doesn't it get in the way?
We were careful about that. The card only shows when there's room beside the conversation, and the conversation moves over to make space. Opening the side panel (the agent's screen, files or work) puts the card away; folding the panel brings it back. If you'd rather not see it, the » in its corner hides it and leaves a small tab with the agent's face. Your choice is remembered on that device.
What keeps this safe?
Nothing about New chat changes what an agent may do. Everything that counts still waits for your OK: sending, paying, deleting, posting in your name. A job that moves to the new chat keeps its own OK cards, and they still need you.
Earlier chats stay as private as any chat. Memory is unchanged and still yours to read, edit or delete, and an incognito task still leaves nothing behind.
And one more thing: the app updates itself
The Mac app and every browser tab show the live nox.cool. Until now, a window left open kept running the version it had loaded, so a change we shipped in the afternoon only reached you after a restart. Now the app checks every few minutes and whenever you come back to it. When there's a newer version, a bar at the top says so, with Update. Unsent words stay in the box. If the window is in the background, it updates by itself after two minutes, but never in the middle of a call.
Try it
nox.cool is open to everyone now, with no invite code. Sign in on the web or download Nox for Mac, ask Nox something, then press the pen button and see what it still remembers. The short version of everything new is in Anyone can sign up, and every agent gets a new chat.
Questions and answers
Does starting a new chat delete my AI agent's memory?
On nox.cool, no. Memory belongs to the agent: the notes it keeps about you, the people you deal with and how you like things done. A new chat starts the conversation over, and the agent still reads those notes. You can read, change or delete any of them on the Memory page.
When should I start a new chat with an AI agent?
When you're moving to an unrelated subject, when its replies keep referring to something you've dropped, when an instruction from earlier keeps getting in the way, or when the chat has run for weeks and you want a clean slate. If the next thing builds on what you just did, stay in the same chat.
Can I go back to an old chat on nox.cool?
Yes. Each earlier chat is listed on the agent's card, named after its first message, newest first. Open one and it's there as you left it; you can keep writing in it, and the agent picks up from there.
What happens to a job that's still running when I start a new chat?
It keeps running and moves to the new chat. Its card shows up there as soon as its status changes, and its result reaches the agent with your next message in the new chat.
Why do long AI chats get worse?
An AI model reads the conversation every time it answers. A very long one costs more to read, takes longer, and mixes old details with new ones, so an instruction from three weeks ago can colour today's answer. Tools like nox.cool summarise long chats to help, but a clean start is often simpler.
Try it with Nox
Anyone can sign up: on the web, or with Nox for Mac.
Keep reading

How should an AI agent show you it's working?
An AI agent can't know how far along a task is. Here's how nox.cool shows which agent is at work, whether it's moving, and when it needs you, honestly.

Why should an AI agent's face show what it's doing?
An AI agent works out of sight. How nox.cool gave each agent a colour you can spot and a moving picture of its job, and took it into the meeting's chat.

What should an AI agent's screen look like?
An AI agent's app has three jobs: take the task, show the work, ask before anything matters. How we rebuilt every screen of nox.cool around those three.