面向多角色人工智能学习伴侣的分层记忆检索与反幻觉方法

A Layered Memory Retrieval and Anti-Hallucination Method for Multi-Persona AI Learning Companions

摘要: 针对多角色学习伴侣场景下记忆共享缺乏隐私边界、检索排序单一、注入记忆易诱发幻觉的问题，提出一种分层记忆检索与反幻觉方法：设计learning/casual/emotional三级可见性模型，实现学习记忆跨角色共享、情绪记忆隐私隔离；提出融合相似度、重要度、时间衰减与模式加权的检索排序，结合低分阈值实现无关查询拒绝；构建记忆注入的反幻觉约束。在408考研场景合成评测集上，方法Recall@5达0.91、正确拒绝全部无关查询(6/6)，跨角色私密记忆泄漏由29~44降至0；加入grounding规则后误引率由40.0%降至13.3%。

Abstract: To address privacy-boundary, ranking, and memory-injection hallucination issues in multi-persona AI learning companions, this paper proposes a layered memory retrieval and anti-hallucination method. A three-level visibility model (learning/casual/emotional) enables cross-persona sharing of learning memories, conditional sharing of casual memories by importance, and privacy isolation of emotional memories. A ranking strategy integrating similarity, importance, time decay, and mode weighting, combined with a low-score threshold, rejects irrelevant queries; anti-hallucination grounding rules suppress injection hallucination. On a synthetic 408 postgraduate benchmark, the method achieves Recall@5 of 0.91, rejects all irrelevant queries (6/6), and reduces cross-persona private leakage from 29~44 to 0. Ablation verifies each module; the grounding rule cuts the memory misquote rate from 40.0% to 13.3%.