发现信息

  • 作者:Farrell Santoso
  • 年份:2026
  • 发表源:DR-NTU (Nanyang Technological University)
  • OpenAlex ID:https://openalex.org/W7163793670
  • DOI:待核验
  • arXiv:待核验
  • 检索主题:agent-memory
  • OpenAlex 引用数快照:0(2026-11-12)

摘要线索

The ability of Large Language Models to perform as autonomous conversational agents is fundamentally constrained by the fixed context window of the transformer architecture. Long-term memory mechanisms that approach memory storage and management in vastly different ways have emerged as essential architectural solutions to maintain cross-session continuity and personalization. This study evaluates the effectiveness of these strategies against current benchmarks, including LoCoMo, LongMemEval, and MemoryAgentBench. Empirical results demonstrate that while hierarchical and graph-based storage architectures provide superior multi-hop reasoning, they face steep engineering costs and the risk of semantic loss during information extraction. Furthermore, models that excel at factual recall often fail in agentic tasks requiring long-range consistency and selective forgetting. Beyond memory recall, evaluation frameworks overlook the open problem of data governance. While functional performance in memory retrieval and recall is maturing, the field lacks standardized methodologies to ensure privacy protection and compliance with the "right to be forgotten" in persistent stores. Future research should expand the development of security-centric benchmarks, such as applying the LLM-PBE benchmark to agents with long-term memory systems, and explore machine unlearning for auditable data deletion. By honing in on improving memory systems according to the most important metrics for an effective conversational agent, the next generation of conversational agents can come one step closer to genuine reliability in real-world deployments.

核验与筛选

  • 当前状态:待筛选、去重和一手来源核验。
  • OpenAlex 元数据和引用数只用于发现,不作为正式引文或质量判断。
  • 记录正式版本、venue、方法细节、实验协议和局限后,才迁入稳定论文卡片。