LLM 智能体何时该遗忘?查询关键事实误删下的记忆过滤研究
When Should LLM Agents Forget? Memory Filtering Under Query-Critical Resolver Errors
AI 导读
一项 arXiv 研究显示,LLM 记忆系统中的硬删除过滤在 resolver 出错并误删查询关键事实时会降低下游准确率。在三个模型上的干预实验中,hard-drop 收益从完美分辨时的 +0.186 至 +0.230 变为最低可靠性下的 -0.176 至 -0.181,且误删造成的退化显著大于误保留。作者建议将删除视为成本敏感决策:在匹配数据上验证误删风险,否则默认无标签保留。
正文 · 原文
Abstract:LLM memory systems often delete facts classified as stale. We show that this hard-filtering step can reduce downstream accuracy when resolver errors delete query-critical facts. Across three models, a controlled sensitivity intervention moves hard-drop from gains of +0.186 to +0.230 at perfect resolution to losses of -0.176 to -0.181 at the lowest tested reliability. Paired targeted interventions show that false-drop causes significantly greater degradation than false-keep, while a non-deleting baseline can still fail when incorrect labels steer generation. On a frozen 102/102 calibration/test split, a simple deployment-matched gate reaches 0.655 accuracy versus 0.641 for always hard-drop and 0.662 for test-informed oracle selection. Global-accuracy crossovers require gold labels and are model- and task-specific diagnostics, not deployment thresholds. Deletion should instead be treated as cost-sensitive: validate query-critical false-drop risk on matched data, or default to label-free retention when such validation is unavailable.
| Comments: | 12 pages, 2 figures |
| Subjects: | Artificial Intelligence (cs.AI) |
| Cite as: | arXiv:2611.17338 [cs.AI] |
| (or arXiv:2611.17338v1 [cs.AI] for this version) | |
| https://doi.org/10.48550/arXiv.2611.17338 arXiv-issued DOI via DataCite |
Submission history
From: Samarth Vinayaka [view email]
[v1]
Sun, 30 Aug 2026 21:52:39 UTC (37 KB)
来源:arXiv cs.AI · arxiv.org