Maglev: Sliding Recurrent Memory

Abstract

We introduce \ours{}, a recurrent Transformer architecture with fixed-size memory that generalizes sliding-window attention while remaining parallelizable during training. \ours{} consists of two coupled models: a prefiller $Q$, which leverages full attention\footnote{In practice, we use interleaved full and sliding-window attention for $Q$, as this yields stronger performance. The essential requirement is that $Q$ be more expressive than $P$, with access to the full history.} to produce memory targets $m_t^{\prime }$, and a decoder $P$, which uses only sliding-window attention and recurrent K/V injection to produce decoder memories $m_t$ for next-token prediction. We train \ours{} with a memory consistency loss that aligns $m_t$ with $m_t^{\prime }$, allowing inference to use $P$ alone. Empirically, \ours{} improves validation loss and downstream pretraining benchmarks over sliding-window and latent recurrent transformer baselines. Moreover, sharing parameters between $P$ and $Q$ reduces parameter memory while preserving most of the gains.