When to Memorize and When to Stop: Gated Recurrent Memory for Long-Context Reasoning

TL;DR

This paper introduces GRU-Mem, a gated recurrent memory model that improves long-context reasoning in large language models by controlling memory updates and exits, leading to better performance and efficiency.

Contribution

The paper proposes a novel gated recurrent memory mechanism with text-controlled gates and reinforcement learning signals to enhance long-context reasoning in LLMs.

Findings

GRU-Mem outperforms MemAgent in reasoning tasks.

Achieves up to 400% faster inference speed.

Demonstrates improved stability and efficiency.

Abstract

While reasoning over long context is crucial for various real-world applications, it remains challenging for large language models (LLMs) as they suffer from performance degradation as the context length grows. Recent work MemAgent has tried to tackle this by processing context chunk-by-chunk in an RNN-like loop and updating a textual memory for final answering. However, this naive recurrent memory update faces two crucial drawbacks: (i) memory can quickly explode because it can update indiscriminately, even on evidence-free chunks; and (ii) the loop lacks an exit mechanism, leading to unnecessary computation after even sufficient evidence is collected. To address these issues, we propose GRU-Mem, which incorporates two text-controlled gates for more stable and efficient long-context reasoning. Specifically, in GRU-Mem, the memory only updates when the update gate is open and the recurrent loop will exit immediately once the exit gate is open. To endow the model with such capabilities, we introduce two reward signals $r^{\text{update}}$ and $r^{\text{exit}}$ within end-to-end RL, rewarding the correct updating and exiting behaviors respectively. Experiments on various long-context reasoning tasks demonstrate the effectiveness and efficiency of GRU-Mem, which generally outperforms the vanilla MemAgent with up to 400\% times inference speed acceleration.