Sessions, memory, and background runs

The left side of the assistant drawer is the session list: one topic per session, no cross-contamination. This page covers how to manage sessions, what long-term memory is, and why a job keeps running after you close the drawer.

Managing sessions

• Click New for an empty session; the title is generated from your first message, and you can rename it from the session's more menu.
• The search box at the top finds past sessions by title and by message content.
• Turn on multi-select to select all and delete in bulk; a single session is deleted from its more menu.
• A whole conversation exports to Markdown in one click, handy for pasting into a ticket or team doc.
• The sidebar collapses, giving the width back to the conversation on small screens.

Long-term memory

Beyond the current session's context, the assistant distils cross-session useful facts into long-term memory — your naming habits, the systems you usually connect, the business conventions you explained last time — so you do not have to repeat yourself.

• Memory is isolated per account; nobody else's assistant can read yours.
• "Clear long-term memory" at the bottom of the sidebar wipes only the semantic memory; the session records themselves stay.
• Clear it once when your business direction changes, or when old conventions from early conversations start working against you.

Runs in the background: closing the drawer does not stop it

A complex task — building a long flow, diagnosing a big run — can take a minute or two. Every assistant request becomes a server-side job that does not depend on this browser window staying alive:

• collapse the drawer and carry on elsewhere in the product — the job keeps running;
• refresh the page or even reopen the browser — returning to this session reattaches to the running job and fills in whatever it produced meanwhile;
• hit stop if you no longer want to wait — the stop request occasionally arrives a little late, and the interface says plainly that the background job may still be wrapping up.

Quotas & Costs

• Conversations that go through the model count against your plan's assistant quota and use the provider and key you configured.
• Slash commands and pure navigation requests call platform endpoints directly, bypass the model and therefore use no quota.
• Asking the same informational question again within a short window reuses the previous answer instead of billing twice.

Read next

• AI assistant overview — What it can do and how to set it up.
• Subscriptions and Pricing — The assistant quota on each plan.
• BYOK: Bring Your Own Key — Use your own key so token costs land on your own account.