Forgetful but Faithful: A Cognitive Memory Architecture and Benchmark for Privacy-Aware Generative Agents (2512.12856v1)

Abstract: As generative agents become increasingly sophisticated and deployed in long-term interactive scenarios, their memory management capabilities emerge as a critical bottleneck for both performance and privacy. Current approaches either maintain unlimited memory stores, leading to computational intractability and privacy concerns, or employ simplistic forgetting mechanisms that compromise agent coherence and functionality. This paper introduces the Memory-Aware Retention Schema (MaRS), a novel framework for human-centered memory management in generative agents, coupled with six theoretically-grounded forgetting policies that balance performance, privacy, and computational efficiency. We present the Forgetful but Faithful Agent (FiFA) benchmark, a comprehensive evaluation framework that assesses agent performance across narrative coherence, goal completion, social recall accuracy, privacy preservation, and cost efficiency. Through extensive experimentation involving 300 evaluation runs across multiple memory budgets and agent configurations, we demonstrate that our hybrid forgetting policy achieves superior performance (composite score: 0.911) while maintaining computational tractability and privacy guarantees. Our work establishes new benchmarks for memory-budgeted agent evaluation and provides practical guidelines for deploying generative agents in resource-constrained, privacy-sensitive environments. The theoretical foundations, implementation framework, and empirical results contribute to the emerging field of human-centered AI by addressing fundamental challenges in agent memory management that directly impact user trust, system scalability, and regulatory compliance.

• The paper presents MaRS, a cognitive memory framework that organizes episodic, semantic, social, and task memories using a graph structure with differential privacy controls.
• It evaluates six forgetting policies, including FIFO, LRU, priority decay, reflective summary, random-drop, and hybrid approaches to optimize memory retention against computational costs.
• The FiFA benchmark demonstrates that hybrid policies yield strong performance, achieving a composite score of around 0.911 while balancing narrative coherence, goal completion, and privacy preservation.

Forgetful but Faithful: A Cognitive Memory Architecture and Benchmark for Privacy-Aware Generative Agents

Introduction

The paper "Forgetful but Faithful: A Cognitive Memory Architecture and Benchmark for Privacy-Aware Generative Agents" (2512.12856) introduces a novel approach to memory management within generative agents that addresses the dual challenges of privacy and computational efficiency. It presents the Memory-Aware Retention Schema (MaRS), a framework inspired by cognitive models which organizes different types of memories—episodic, semantic, social, and task-oriented—as discrete entities within a graph structure. This architecture is designed to ensure efficient retrieval and principled forgetting, supported by various policy mechanisms that include differential privacy guarantees.

Memory-Aware Retention Schema (MaRS)

MaRS is the centerpiece of the study, conceptualized as a structure where memories are nodes with metadata tracking provenance, sensitivity, and computational weight. The infrastructure categorizes memories into distinct types: episodic, semantic, social, and task memories. This segregation allows the system to apply targeted retention and forgetting strategies that align with both cognitive relevance and privacy norms. Additionally, MaRS supports various forgetting policies—FIFO, LRU, priority decay, reflective summary, and hybrid approaches—that balance memory retention against computational cost and privacy constraints.

Forgetting Policies

Six forgetting policies are formalized within the MaRS framework, each offering different trade-offs between computational complexity and memory retention fidelity:

• FIFO and LRU: These are temporal heuristics focusing on recency, where FIFO operates by evicting the oldest memories and LRU by removing least-recently-used items.
• Priority Decay: This approach ranks memories by a calculated importance score, prioritizing retention of high-utility items.
• Reflection-Summary: This method consolidates related episodes into semantic summaries to reduce storage requirements while maintaining narrative coherence.
• Random-Drop and Hybrid Approaches: These incorporate elements of randomness or multiple strategies to address varying demands on the memory system.

Each policy's effectiveness is analyzed concerning different operational scenarios and memory budget constraints, with particular emphasis on privacy-aware mechanisms that include optional differential privacy settings.

Evaluation with FiFA Benchmark

To evaluate these strategies, the FiFA benchmark was introduced. It assesses agent performance across multiple axes: narrative coherence, goal completion, social recall, privacy preservation, and cost efficiency. The benchmark emphasizes real-world interaction scenarios where agents engage over extended sessions, providing a rigorous testbed for these memory strategies. The results showed that hybrid policies generally offered the best combination of performance and efficiency, achieving a composite score around 0.911 under simulation conditions.

Implications and Future Directions

The research highlights the potential of MaRS not only in enhancing the operational efficiency and privacy compliance of generative agents but also in aligning their memory processes more closely with human cognitive paradigms. This work paves the way for future developments in AI systems that require a nuanced understanding of information retention, particularly in contexts where privacy and efficiency are paramount. Future work could explore more adaptive forgetting strategies, potentially incorporating more sophisticated machine learning techniques to further optimize memory management in dynamic environments.

Conclusion

The "Forgetful but Faithful" framework advances the design of generative agents by integrating cognitive principles with technical mechanisms for memory management. The combination of the MaRS architecture and the FiFA benchmark offers a comprehensive approach for developing privacy-aware AI systems, emphasizing the critical role of memory governance in AI. This work sets a foundation for deploying cognitive agents capable of maintaining operational integrity while respecting privacy constraints, thereby enhancing their trustworthiness and efficacy in real-world applications.