Keep a long conversation in bounds

A long conversation eventually fills the model’s window. From the conversation, open its context panel: it measures what the chat uses, prepares a summary when needed, and lets you decide what must not be lost. Why it matters is explained in What the model reads.

Let TALOS manage it

1. Turn on automatic management.
2. In the advanced settings, set the percentage of the window at which compaction starts, and the one it brings the context down to. The target must be lower than the start; the opposite is refused, with the reason.
3. Choose the model that writes the summary: the chat’s own, or another one.

Compact by hand

• Compact now makes a summary at once.
• Regenerate the summary when the one there does not convince you.

A compaction goes through named steps — waiting, preparing, compacting, checking the summary, publishing, updated — and, when something goes wrong, it says “compaction paused”, “compaction canceled” or “compaction failed”.

Keep what matters

Before compacting, add to Facts to keep what must survive the summary. Those facts stay separate from it, and you can add, edit and remove them. When TALOS proposes to change a protected fact, the proposal waits for you: accept it, or keep the fact.

Go back

Versions lists the saved summaries. Restore an earlier one; the messages after it stay in the chat. The originals are never deleted: the summary is what the model reads, not what stays on disk. It applies to this chat only.

If something goes wrong

• The context changed — someone, or the model, touched the context while you were editing. Read it again and repeat the change.
• The context is not available — the reading did not arrive. Refresh to try again.
• Compaction failed — no summary was made and the originals are intact. Try again, or compact by hand.
• The measure is missing — it says so (“not available yet”) rather than showing a zero. Native compaction is not offered for every model; where it is not qualified, the row says so.