Agent conversations

Inbound user turns from the Hermes and OpenClaw integrations archive like any other stream — under the agent stream, threaded by session. They matter twice: they are evidence ("Poker moved to Saturday, I'm still going" is a genuine user correction), and they are the turns freshness hints must reach.

Only the user's original message is kept as user-authored evidence. Assistant replies, generated summaries, injected snapshots, and tool output are excluded on the archival path — never ingested as if they were user facts. That boundary is what keeps a wrong assistant answer from reinforcing itself at the next dream: the original source evidence and your actual words decide, not the assistant's Saturday answer.

A user asking "Is poker still Saturday?" is asking, not affirming Saturday. "You told me Saturday" attributes a claim to the assistant; it is not an independent correction. Both cases run through the ordinary semantic review with roles and attribution preserved. See Daytime freshness and Evidence & provenance.