Memory Layer: Make the recommendation cache part of model training¶
Fidelity: core mechanism reproduction. Executes in-training writeback, shared cache reads, a frozen external snapshot control and always-on miss fallback. This is not Meta's MPZCH infrastructure.
Paper information¶
| Field | Details |
|---|---|
| Paper | Paper or original source |
| Institution | Meta / Instagram Reels |
| First public release | 2026-07-27 (arXiv v1) |
| Author code | Yes for shared infrastructure: TorchRec, with MPZCH, Writeback and RES components in TorchRec/FBGEMM. Full model integration is internal. |
| Local adapter or method key | memory-layer |
| Local reproduction code | src/auto_research/reproductions/memory_layer/ |
Paper method and evidence¶
Early-stage rankers often train with fresh item-tower features but serve a separate precomputed cache; misses may bypass scoring, so training/evaluation/serving see different inputs. Memory Layer makes the cache a sparse model table: the tower writes during training with learning rate one, and downstream and serving read the same representation. A cache miss uses a zero embedding plus always-on attributes such as creator/topic.
flowchart LR
A[Item tower] --> B[Training writeback]
B[Training writeback] --> C[Shared cache read]
C[Shared cache read] --> D[Always-on fallback]
D[Always-on fallback] --> E[Scoring]Original paper figure¶
Figure 1 shows writeback and the shared feature flow. From the original paper; copyright remains with the authors.
Core equation¶
The tower outputs \(e_i=f_{item}(x_i;\theta_t)\); scoring uses \(\hat e_i\) and generalizable attributes \(a_i\).
Reported results¶
Instagram Reels coverage goes 96%→100%, freshness from about five minutes to about 20 seconds, and pselect training/serving NE gap 12.11%→1.64% (86% relative reduction). A/B backtest: plays of media under one hour rise 6%–7%, under five minutes more than 2×, and reshares/time spent 5%–6%. Training and publishing compute fall 30%, with neutral serving compute.
Local reproduction¶
Matched baseline uses the same tower with 96% external frozen cache capacity and no always-on fallback; treatment writes back during training and can score 100% of items. Seed 42 NDCG@10 improves about 4.36% relative; structural diagnosis is the main point, not public-proxy lift. Processes training events in order with eta=1, logging coverage, mean staleness steps and writeback error. Three-seed metrics, SD and 95% CI.
MovieLens lacks creation time, Instagram cold-start exposure and online NE. No distributed MPZCH/FBGEMM store, 15-second streaming, int8 row-wise quantization or production cache capacity.