Decoding Post-Search Behavior with PALACE: Role-Structured Behavioral Memory and Evidence-Isolated Multi-Agent Reasoning

Decoding what happens after search requires more than predicting the next item: from pre-cutoff history alone, a model must determine whether a search-heavy shift develops into target continuation, ends in ocal resolution, returns to a prior category, or remains unresolved. These trajectories can share similar recent queries yet differ in sparse search–consumption contrasts, missing temporal bridges, and later reversals—signals that are easily buried in redundant logs. Existing LLM-based methods flatten such logs or let multiple agents inherit the same intermediate judgment, causing frequent but weak events to dominate decisive exceptions and correlated interpretations to masquerade as consensus. We introduce PALACE, a framework that addresses these two failure modes through two complementary innovations. First, role-structured behavioral memory replaces generic recency- or relevance-based summarization with hypothesis-differentiating compression: it merges repetition into source-linked anchors, preserves contradictions and temporal turns, and organizes each trace by what it can establish or refute. This protects low-frequency but trajectory-decisive evidence while retaining provenance. Second, evidence-isolated multi-agent reasoning separates candidate generation and falsification from a verdict-blind audit of behavioral continuity, then arbitrates the independently constrained records. This design reduces shared-framing bias and yields a traceable prediction without task-specific parameter updates, retrieved cases, or access to future behavior. Comprehensive experiments demonstrate that PALACE consistently outperforms strong LLM-based prompting, memory, and debate baselines in predicting post-search trajectories. Further analyses confirm that organizing behavioral memory by evidential role and isolating complementary agent judgments are critical to these gains, while supporting robust transfer across behavioral schemas and observation windows.