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.
Large language models (LLMs) operating as generative agents are increasingly deployed in open-ended interactions that span hours or days, blending dialogue, planning, tool use, and reflection across sessions (Park, O'Brien, Cai, Morris, Liang and Bernstein, 2023;Wang, Ma, Feng, Zhang, Yang, Zhang, Chen, Tang, Chen, Lin et al., 2024). In such settings, the capacity to remember and forget becomes a first-class design variable rather than a peripheral implementation detail. The problem resembles long-standing questions in human memory research: agents must organize experiences into episodic traces, consolidate them into semantic knowledge, track social relationships, and preserve task context under bounded resources (Baddeley, Aggleton and Conway, 2002;Tulving, 2002). Unlike traditional stateless systems, however, contemporary agents continuously accrue interaction history. Unchecked growth stresses inference time through long contexts, increases retrieval noise, and amplifies privacy risk by retaining sensitive details beyond any reasonable need (Carlini, Tramer, Wallace, Jagielski, Herbert-Voss, Lee, Roberts, Brown, Song, Erlingsson et al., 2021;Brown, Lee, Mireshghallah, Shokri and Tramer, 2022).
Cognitive architectures have long separated episodic, semantic, and procedural knowledge and introduced activationor recency-based decay to manage capacity (Anderson, 1996). Modern agent frameworks have revived these ideas in neural settings, adding reflective summarization and retrieval-augmented generation to stabilize behavior over long horizons (Schank and Abelson, 1977;Shinn, Cassano, Gopinath, Narasimhan and Yao, 2023). Yet current practice often oscillates between two extremes. At one end, systems defer deletion entirely, relying on ever-longer contexts and vector stores; this improves short-term recall but degrades latency and raises privacy and governance concerns. At the other, ad-hoc pruning (e.g., fixed windows, random drops) preserves efficiency at the cost of narrative coherence, goal continuity, and social appropriateness (Liu, Yu, Zhang, Xu, Lei, Lai, Gu, Ding, Men, Yang, Zhang, Deng, Zeng, Du, Zhang, Shen, Zhang, Su, Sun, Huang, Dong and Tang, 2024). Neither approach offers principled guarantees about what is retained, what is forgotten, and why. This paper advances forgetting-by-design as a human-centered principle for generative agents: memory should be structured, budgeted, and explicable, with retention decisions aligned to task value and privacy norms. We introduce the Memory-Aware Retention Schema (MaRS), an ontological and operational layer that represents memories as typed nodes with provenance, sensitivity, and token-weight metadata, linked by relations that support efficient retrieval and principled removal. MaRS is paired with a palette of forgetting policies that range from temporal heuristics (FIFO, LRU) to importance-aware methods (priority decay), reflective consolidation (summary-based compression), and a hybrid scheme that stages these mechanisms to balance fidelity with cost. Crucially, retention decisions can be modulated by a privacy engine that supports sensitivity weighting and optional differentially private noise injection, bringing the practice closer to emerging expectations around transparency, accountability, and the right to be forgotten (Amershi, Weld, Vorvoreanu, Fourney, Nushi, Collisson, Suh, Iqbal, Bennett, Inkpen et al., 2019;Jobin, Ienca and Vayena, 2019).
Evaluating memory management requires metrics that capture more than task success. We therefore propose the Forgetful but Faithful Agent (FiFA) benchmark, which measures narrative coherence across turns, completion of multi-step goals, accuracy of social recall, privacy leakage, and cost efficiency under explicit token budgets. FiFA is implemented as a multi-agent simulation with controllable pressures on memory growth and retrieval selectivity, allowing fair comparisons across policies and budgets. In contrast to single-task or purely offline evaluations, FiFA targets the lived properties of agent interaction-maintaining continuity with a user while avoiding unnecessary retention of sensitive details.
Our empirical study shows that purposeful forgetting can improve both user-facing quality and governance posture. Across extensive runs and budget settings, hybrid policy variants consistently preserve coherence and social recall while keeping leakage low and costs tractable, whereas naive strategies either overspend on context or erode interaction quality. Beyond aggregate scores, MaRS’s audit and provenance traces render retention choices interpretable post hoc, a prerequisite for trustworthy deployment in domains that demand accountability.
The contributions of this work are conceptual, algorithmic, and evaluative. Conceptually, MaRS frames memory as a relational, provenance-aware store with explicit budgets and privacy semantics. Algorithmically, we design and analyze a family of forgetting policies, incl
This content is AI-processed based on open access ArXiv data.