Descendant Closure Knockout for Context Management in Long-Horizon Search Agents

Long-horizon search agents accumulate tool calls and web observations over extended trajectories, causing append-only histories to mix useful evidence with redundant content and eventually exhaust the context window. In existing long-horizon search agents, retention decisions are typically made by the summarizer or the search policy itself, without a separate observation-level predictor trained from gold-answer-conditioned influence. A natural way to construct such a signal is single-observation knockout, but it can underestimate delayed influence because later queries may reuse entities introduced by early evidence, leaving downstream observations that continue to carry information from the root in the counterfactual context. We introduce Descendant Closure Knockout (DCK), a lightweight retention layer that ranks observations for protected compaction while keeping the search agent, compaction trigger, and summarizer frozen. Offline, DCK constructs a temporal entity-lineage proxy from entity reuse in later queries and jointly masks a root observation with its currently addressable lineage descendants. It measures the change in the frozen agent’s teacher-forced gold-answer log-likelihood and distills this horizon-conditioned closure-knockout influence into an agent-specific lightweight head. At deployment, when the host triggers compaction, DCK preserves bounded verbatim blocks from the highest-scoring observations within the fixed reset-context budget, while the host summarizes the remaining history. We evaluate DCK across frozen search agents of different scales, multiple summary tools, with two context-compaction instantiations, using matched context and tool-call budgets within each paired comparison.