This study focuses on the tendency of agents in long-horizon sequential tasks to rely on short-term states and to underutilize historical information, and proposes a cognitive modeling and learning framework with long-term memory and reasoning capabilities. The framework provides a unified cognitive description of the agent's decision process. It introduces a structured long-term memory mechanism to support continuous storage and selective updating of cross-temporal key information. On this basis, a memory retrieval-driven reasoning module is constructed so that experience can explicitly participate in the formation of current decision logic. To address the separation between memory and decision making in conventional policy models, the framework tightly couples perception representation, memory management, reasoning processes, and policy generation into an end-to-end cognitive loop. This design strengthens goal consistency and behavioral stability in long-horizon interactive environments. Comparative evaluations in open source interactive task settings demonstrate consistent advantages in task completion quality, decision efficiency, and long-term information utilization. The results indicate that the proposed cognitive modeling framework effectively mitigates decision difficulties caused by long-range dependencies and partial observability. Overall, the study shows that integrating long-term memory and reasoning within a unified learning framework is an important approach for improving sustained decision-making capability in complex environments.

Large Language Model (LLM)-based multi-agent systems have demonstrated strong capabilities in collaborative task-solving. However, a practical challenge emerges in extended collaboration: role drift, where agents gradually deviate from their designated responsibilities. This phenomenon manifests as boundary violations (e.g., a planner writing code), redundant work, conflicting decisions, and futile debates, ultimately degrading system performance. In this paper, we present RoleFix, a lightweight framework for detecting and repairing role drift in multi-agent collaboration. Our approach introduces: (1) a structured protocol requiring agents to declare their role, commitments, and dependencies at each turn; (2) a hybrid drift detector combining rule-based checks with LLM-based semantic judgment; and (3) a self-repair mechanism inspired by verbal reinforcement learning that triggers reflection, role reassignment, and execution resumption. Experiments on software engineering and research workflow tasks demonstrate that RoleFix reduces role drift incidents by 67.4% and improves task completion rates by 23.8 percentage points compared to baseline multi-agent systems, while introducing only 8.3% latency overhead.

Large language model (LLM)-based multi-agent systems have demonstrated remarkable capabilities in collaborative task solving. Although the mechanisms that facilitate seamless cooperation, such as shared contexts, role assignments, and iterative message passing, present significant risks of unintentional information disclosure. We present MIN-Trust, a trust orchestration framework that enforces Minimum Necessary Information (MNI) constraints, an operationalization of the data minimization principle for inter-agent communication—while maintaining task effectiveness. Our approach introduces an MNI-Gate that automatically classifies and filters information into essential, summarized, or pointer-referenced subsets before transmission. Additionally, we propose a Trust-Gated Channel (TGC) that counterintuitively increases verification requirements rather than relaxing information access as inter-agent trust elevates. Through experiments on four collaborative tasks using public benchmarks, we demonstrate that MIN-Trust reduces sensitive information exposure by 67.8% compared to baseline multi-agent frameworks while maintaining 93.3% of task success rates. Our evidence traceability mechanism achieves 84.2% claim-to-source attribution, significantly outperforming conventional approaches. These results suggest that privacy-preserving multi-agent collaboration is achievable under synthetic benchmark conditions with moderate performance trade-offs.