Drift-Aware Byte-Utility Placement for Tiered Embedding Memories in Recommendation Models

The memory footprint of categorical embeddings is often the dominant cost of neural recommendation. Frequency-aware compression and static placement exploit skew, but a placement chosen from historical frequencies can become inefficient as popular IDs change. We study an online placement extension of a tiered hot/warm/cold embedding layer that evaluates saved slow-memory bytes rather than only cache-hit mass and penalizes row migration. We implement full-dimensional hot rows, projected low-dimensional warm rows, and quotient–remainder cold rows in an executable recommendation model. In a fully specified synthetic dataset of 128,000 chronologically ordered user–item events, three training seeds yield a mean test AUC of 0.6842 at 16× item-table compression, compared with 0.6854 for a full item table. Trace-driven replay at 8× and a fixed cache capacity gives 27.37 slow bytes/request for static access-density placement and 25.73 for the churn-aware controller, including migration traffic. An exact LRU cache remains stronger on this workload (17.33 bytes/request). Sixteen independent synthetic traces per drift regime characterize stability across stationary, abrupt, gradual, and cyclic popularity, with significant variation in benefit. The study isolates a testable byte-utility placement mechanism but does not substitute for public-data or measured-hardware validation.