ConeScope: Learning When Procedural Memories Apply

摘要

Procedural memories enable long-horizon agents to reuse previously acquired workflows and skills. Existing memory systems primarily focus on memory construction and semantic retrieval, while leaving a central question underexplored: when should a retrieved procedural memory actually be used? Semantic relevance alone does not imply procedural applicability. As execution progresses, the same memory may become beneficial, irrelevant, or harmful, making memory utilization a state-dependent control problem rather than merely a retrieval problem. We formulate procedural-memory applicability through the downstream effect of using a memory at the current execution state. To obtain direct supervision, we construct exact-prefix paired rollouts that compare a memory intervention with its reference action from the same environment state, interaction history, and active-memory status. We then propose ConeScope, which embeds memories and observable execution states in a Poincaré ball and uses hyperbolic entailment cones to model the directed applicability relation from a memory to execution states. Cone violation, radial order, semantic information, and execution features are jointly used to predict both the mean effect and a conservative lower-quantile effect. At inference time, the learned geometric scope and lower-quantile prediction jointly govern Enter, Abstain, Keep, and Exit, separating candidate retrieval from state-dependent memory use. Experiments on ALFWorld, ScienceWorld, and τ²-Bench with 7B and 32B action backbones show consistent improvements over retrieval- and memory-control baselines. Controlled component and geometry studies further support the contributions of selective intervention, lifecycle review, conservative effect prediction, and directed scope modeling. These results support treating procedural memories as state-dependent control objects with learnable applicability scopes, rather than as static retrieved knowledge.