The live Python session
Each chat has its own Python session that keeps running between messages. Variables, imports, and helper functions the model creates stay available, so a follow-up can use them straight away instead of starting over.
A session lasts until the app closes, the session is restarted or released, or something crashes. Sessions don’t time out on their own. Overview → Python shows every live session, what it is running right now, and how much CPU and memory it uses, with a button to release one when you are done.
What lasts, and for how long
| What | Holds | Lasts |
|---|---|---|
| Chat history | The conversation and its records. | Saved. It describes past work, but doesn’t keep Python objects alive. |
| Live session | Variables, imports, helpers, open connections. | Until the session restarts or is released. |
| Files | Anything written to disk. | Until you delete them. Check they were actually saved. |
| Checkpoints | A copy of supported values. | Optional and off by default. Not everything can be copied. |
| Session context | An index of the chat’s own history and results. | Saved with the chat, so the agent can look things up later. |
Optional checkpoints
Checkpoints save supported Python values so they can be brought back after a restart. Both saving and restoring are off by default.
A checkpoint is not a full copy of a running session. Open files, browser connections, and background threads have to be set up again, and a saved helper may depend on data that is gone.
Looking back with session context
Agents can search and reread a chat’s earlier messages, code, and results when they need them. In Settings → Session context you can browse the same evidence, search it, see where each piece came from, and export it as JSONL or Markdown.
There is no hidden memory. VARIANT-1 doesn’t quietly build a profile of you or carry facts from one chat into another. Memory across chats is not part of this preview.
Picking up after a break
- Ask whether the session restarted since you last worked in this chat.
- Have the model check the values or files the next step needs.
- Trust what is actually present over what the chat says used to be there.
- If something is missing, decide together how to rebuild it.