Cold-Start Ranking: Architectural Semantic Labels Predict Memory Use Where Graph Methods Go Blind

Abstract摘要

In agent memory systems, graph-based ranking methods go blind at cold-start: 91% of new nodes remain invisible to PageRank and access-count scores throughout the observation window. Architectural semantic labels assigned at node creation predict future organic use at AUC 0.84, a 34-point advantage over random, of which approximately 18 points exceed pure-text baselines (0.66). We document this in a self-built agent research testbed in operation since March 2026. Cold-start is persistent: 91% of nodes created after an August 2026 architectural transition never accumulate a single access (as of October 2026, seven weeks later). The most trivial baseline, newest-first, performs below random (AUC 0.41) because use accrues with time, not against it. We show that the advantage is architectural—creation-time categorical assignments (semantic domain, content category, provenance type) carry signal that shallow lexical features cannot extract. With accumulated history, structural reinforcement takes over (AUC 0.917). Whether the label prior adds increment in this regime was not tested (no within-regime comparison was run); the claim is limited to the cold-start domain. We contribute: a quantified cold-start blind spot, a demonstration that architectural labels serve as cold-start priors, a disentanglement of label vs. text contribution, and a stated boundary (the advantage is regime-dependent and partially circular). No cross-system superiority is claimed.