Agent state consolidation

An assistant runs for one account across months: thousands of sessions, tool calls, corrections and standing instructions. Two things are fixed. The weights are frozen, so no fine-tuning, no adapter, no learned component that needs gradient updates on this account's data. And a retention policy deletes raw events after a rolling window, so the log is not there to search later. What persists between sessions is a consolidated state of at most B bytes, where B does not grow with the age of the account. A consolidation operator runs offline, sees events in one streaming pass, and rewrites that state.

The queries that decide whether this works are the ones the current session gives you no handle on. A new instruction quietly contradicts a preference set four months ago. A plan the agent is about to propose repeats an approach that failed twice before, described in different words both times. A claim it is about to make was corrected by the user in a conversation that shares almost no vocabulary with the claim. Similarity search finds none of these, because nothing in the live turn is close to what you need, and by the time the agent knows to look the raw events are gone.

Hand in the consolidation operator: what it writes, what it drops, and how the state is read at session start under a bounded read budget. State the class of unanticipated queries you claim to serve and the class you are abandoning, in words specific enough that someone could construct a counterexample. Give the experiment that would show your state preserves contradiction-relevant structure rather than preserving whatever the summarizer found interesting.

Ties break toward designs that give a degradation curve as history over B grows, and toward those that say what happens when the consolidation operator itself was wrong six weeks ago.