Small Memories Steer Big Ones: Compact Revision Dependencies for Personalized Memory Targeting

When a user's circumstances change, a long-term memory system must decide not only how to update stored information, but which existing memories are worth reconsidering in the first place. We study this revision-targeting problem through two complementary contributions. First, we introduce REVSCOPE-100, a controlled benchmark of 50 matched counterfactual-history pairs (100 cases): persona and event text are byte-identical within each pair, and the same 30 memory texts appear under case-local permutation, while revision history differs. Thus the active revision dependency has no systematic signal in the current state alone. Second, we introduce UPSCOPE, a compact auxiliary memory that consolidates past revision consequences into reusable trigger->dependency rules and uses them to steer targeting over a much larger factual memory without replaying the full revision history at runtime. Across Qwen3.8-Max and DeepSeek-V4-Pro, raw revision experience substantially improves personalized targeting over current-state reasoning alone. UPSCOPE compresses revision memory by roughly 78% on both backbones (about a 31% reduction in total runtime prompt tokens) while retaining substantial targeting utility. Most notably, replacing only UPSCOPE with one induced from the matched counterfactual history reverses targeting toward the partner targets on both backbones. On the primary Qwen backbone, strong dense retrieval is competitive in targeting accuracy, suggesting that UPSCOPE's main advantage is not universal accuracy gains over retrieval, but a compact, structured, and causally controllable representation of revision experience.