Legal intelligence aims to support reliable decision‑making across long‑horizon processes where case states evolve and multiple roles interact. In real deployments, cases vary widely in facts, evidence, and procedural context, exposing the limits of static agent strategies. Moreover, legal reasoning is highly interdependent across roles and procedural stages, making overall reliability fundamentally different from isolated role competence. To tackle these issues, we explore training‑free test‑time agent adaptation: agents continuously exploit signals from preceding cases and ongoing interactions during deployment without updating model parameters. We introduce \method, which comprises two key components:
- Test‑Time Memory Evolution – retrieves reusable experience from past cases, adapts it to the current factual and procedural context, and consolidates the adapted experience back into memory for future decisions.
- Rubric‑Aligned Collaboration – verifies and revises role‑specific actions according to behavioral and procedural rubrics, enabling coordinated decision‑making across roles and stages.
Extensive experiments on J1‑EVAL and LegalWorld using five backbone models show that \method consistently outperforms representative reasoning and agent baselines while keeping interaction and computational costs reasonable. Ablation studies and case analyses reveal that the two components provide complementary benefits in experience adaptation and cross‑role coordination, markedly improving the reliability and efficiency of long‑horizon legal reasoning.
Review: By dynamically leveraging deployment‑time signals and structured collaboration, \method overcomes the rigidity of traditional static models in legal settings, offering a practical pathway toward more robust and scalable legal AI systems.