Research map / Large language models / Reasoning and chain of thought

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

ICML 2026 · Large language models › Reasoning and chain of thought · approach: Efficient LLM serving · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes GRU-Mem, which adds text-controlled update and exit gates to MemAgent's recurrent chunk-by-chunk memory and trains them end-to-end with RL, reducing memory explosion and inference time for long-context QA.

Problem

Recurrent textual-memory approaches for long-context reasoning (MemAgent) indiscriminately update memory on evidence-free chunks, causing memory explosion, and lack an early-exit mechanism, wasting computation.

Approach

Augment the memory agent's output with an update gate (whether to overwrite memory with a candidate) and an exit gate (whether to stop reading), trained with separate update and exit rewards plus disentangled trajectory-level and turn-level advantages in multi-turn RL.

Key results

• GRU-Mem (w/o EG) at 7B averages better than MemAgent on most tasks with roughly 200% inference acceleration.
• GRU-Mem (w EG) at 7B reaches up to 400% speedup (e.g., MK-1) with comparable or better performance than MemAgent.
• Under evidence-at-top-20% setting, GRU-Mem (w EG) reduces inference time to about 1/4 of MemAgent.

Surprising findings

• Without an update reward (alpha=1), update accuracy on evidence-free chunks drops sharply, indicating indiscriminate updating.

Limitations

• Limited to QA tasks; summarization and other tasks unexplored.
• Extra gate rewards reduce training stability, requiring smaller off-policy degree and longer convergence.

Experimental setup

Datasets: MemAgent training data HotpotQA SQuAD RULER

Benchmarks: HotpotQA SQuAD RULER NIAH (SK-1/2/3, MK-1/2/3, MQ, MV)

Metrics: accuracy (Perf. %) inference time (s)

Baselines: MemAgent

Builds on: MemAgent GRU GRPO DAPO

Closest papers by meaning

• MEM1: Learning to Synergize Memory and Reasoning for Efficient Long-Horizon Agents (ICLR 2026 · similarity 0.88)
• Longer Context, Deeper Thinking: Uncovering the Role of Long-Context Ability in Reasoning (NeurIPS 2025 · similarity 0.87)
• M+: Extending MemoryLLM with Scalable Long-Term Memory (ICML 2025 · similarity 0.87)
• A Little Goes a Long Way: Efficient Long Context Training and Inference with Partial Contexts (ICLR 2025 · similarity 0.87)
• Reasoning Under 1 Billion: Memory-Augmented Reinforcement Learning for Large Language Models (NeurIPS 2026 · similarity 0.87)
• InftyThink: Breaking the Length Limits of Long-Context Reasoning in Large Language Models (ICLR 2026 · similarity 0.87)