Fortunate Recall：基于行为本体的大模型长效记忆生命周期管理

My Thoughts on Managing LLM Memory

I've been examining the critical issue of lifecycle management within Large Language Model (LLM) memory systems. My assessment reveals that current systems treat all personal facts identically, leading to unchecked storage growth and declining retrieval precision. The core challenge lies in the effective lifecycle management of these facts: determining which memories to preserve, which to replace, and at what rate, all conditioned on the behavioral type of each fact. To address this, I've developed Fortunate Recall (FR), a composable policy layer that categorizes personal facts using a "10+1" behavioral ontology. This layer then applies category-specific lifecycle policies – differential temporal decay, slot-key supersession, event-time validity, and category-aware retrieval routing – as deterministic functions operating on metadata extracted by the LLM. My infrastructure-independent implementation, FR-Bank, demonstrates a 76.9% pass rate on the new 516-question temporal-disambiguation benchmark, LifecycleBench, surpassing Mem0, A-MEM, Memory-R1, and MemoryOS (61% to 70.5%). Furthermore, it achieves 75.2% on the full LongMemEval-S under the canonical Wu et al. judge protocol, which proves that lifecycle policies don't introduce a measurable cost on standard retrieval. Through a pre-registered ablation study, I found that replacing the typed layer with three generic lifecycle primitives maintains correctness (-1.7 percentage points, 95% CI [-6.0, +2.7]), indicating that the generic lifecycle metadata provides the correctness benefit, while the behavioral ontology provides calibration, halving downstream confabulation (12.0% versus 24.2%, p < 0.001). End-to-end evaluation reveals that FR-Bank reduces confabulation from Mem0's 45.1% to 22.4% on answered queries and from 32.2% to 13.0% on all queries, while simultaneously improving correct answer rates (31.2% versus 18.6%). This performance advantage is replicated on the open-weight Kimi K2.5 model. This decomposition also transfers to the independently constructed BEAM benchmark, with a 46.8% correct rate over 280 questions compared to Mem0's 32.9%. The benefit of the ontology is particularly evident in contradiction resolution and plateaus at approximately seven policy clusters. Finally, I’ve made the ontology, benchmark, and code publicly available.

当前的 LLM 记忆系统对所有个人事实均同等对待，导致存储规模无限膨胀，同时检索精度不断下降。其核心挑战在于生命周期管理：以每个事实的行为类型为条件，决定哪些记忆应当保留、哪些应当更替，以及更替的速率。Fortunate Recall (FR) 是一个可组合的策略层，它将个人事实归入一套“10+1”行为本体中，并将类别特定的生命周期策略（差异化时间衰减、槽位键替换、事件时间有效性和类别感知检索路由）作为基于 LLM 提取元数据的确定性函数加以执行。作为与基础设施无关的实现，FR-Bank 在包含 516 个问题的新型时间消歧基准 LifecycleBench 上取得了 76.9% 的通过率，优于 Mem0、A-MEM、Memory-R1 和 MemoryOS（61% 至 70.5%）；同时在 Wu 等人提出的权威评判协议下，于完整的 LongMemEval-S 上达到了 75.2% 的通过率，表明生命周期策略并未对标准检索性能造成可测量的损耗。一项预注册消融实验明确了性能提升的来源：用三个通用生命周期原语替代类型化层后，正确率在统计上无显著变化（-1.7 个百分点，95% CI [-6.0, +2.7]），这表明通用生命周期元数据贡献了正确率增益，而行为本体则提供了校准能力，使下游臆造率减半（12.0% 对比 24.2%，p < 0.001）。在端到端评估中，FR-Bank 在已回答的查询中将臆造率从 Mem0 的 45.1% 降至 22.4%，在全部查询中从 32.2% 降至 13.0%，同时正确回答了更多问题（31.2% 对比 18.6%）；该排名结果在开源权重的 Kimi K2.5 上同样得到复现。这种解耦方法还可迁移至独立构建的基准 BEAM：在 280 个问题上取得了 46.8% 的正确率（Mem0 为 32.9%），且本体的优势集中在矛盾消解上，并在约七个策略簇处趋于饱和。相关本体、基准测试集及代码均已开源。