Agents & Cognition · 06.04
Sessions: Persistent Working Memory, Reconstructed Prompts
Sessions persist across turns, but the prompt is reconstructed each time — persona, task, plan, and fresh observations reassembled, not resent stale.
Persistent state, rebuilt prompt
A session survives across many turns — it can park on await_event and wake later, and it survives a full process restart. But what gets sent to the model each turn is never "the whole history plus one more message." Persona, the active task, the current plan, and fresh observations are reconstructed fresh every time, which is what keeps a long-running session from slowly accumulating stale, contradictory context the way a naive append-only transcript would.
Continue
How each task is classified and routed to the cheapest sufficient model tier, unless an explicit pin or override wins.
Inside a run, an agent node wields its full Agentis toolbox — role manifest, universal floor, bridged MCP tools, and the agentis.* surface — as one streamed activity feed.