MARCH Scales Recurrent Memory for Long-Context LLMs Efficiently

Ming Zhang, Kaisen Yang, Shu Yu, Ermo Hua, Ning Ding, Xia Hu, Bowen Zhou, Chaochao Lu, Youbang Sun· August 14, 2026 View original

Key takeaways

  • Transformers have strong long-context retrieval but incur quadratic computation and linear memory growth.
  • Recurrent models are efficient but struggle with long-range recall due to fixed-size states.
  • MARCH scales recurrent memory by caching content-routed state anchors, forming a memory bank.
  • It outperforms linear attention variants on commonsense reasoning and long-context tasks.

Who benefits

AI DevelopmentNatural Language ProcessingSoftware DevelopmentResearch & Development

Summary

MARCH (Memory-Anchor Routing across Context History) is a new network architecture that scales state-space models beyond fixed-size dimensions while maintaining computational efficiency for long sequences. It achieves this by periodically caching recurrent-state checkpoints as content-conditioned state anchors, forming a memory bank for long-range recall.

Transformers excel at long-context retrieval due to their token-level memory, but this comes at the cost of quadratic computation during training and linear memory growth during inference. Recurrent neural networks offer efficient decoding with fixed-size states but often struggle with long-range recall as older information is overwritten. This research introduces MARCH (Memory-Anchor Routing across Context History), an architecture designed to overcome these limitations. MARCH periodically saves cumulative recurrent-state checkpoints as "state anchors," each associated with a compact, content-conditioned key. This creates a memory bank that can expand with context length, offering a controllable balance between historical detail and memory usage. During processing, MARCH queries these historical anchors, aggregating their information with the current state to enhance long-range memory.

Why it matters

Professionals working with large language models can leverage this innovation to build models that handle much longer contexts more efficiently, leading to improved performance on tasks requiring extensive memory without prohibitive computational costs.

How to implement this in your domain

  1. 1Investigate integrating MARCH's content-routed state anchoring into custom recurrent neural network architectures.
  2. 2Benchmark MARCH against existing Transformer and linear attention models for long-context tasks.
  3. 3Explore the trade-offs between historical resolution and memory cost for specific application requirements.
  4. 4Train AI engineers on advanced memory management techniques for sequence models.

Original post by Ming Zhang, Kaisen Yang, Shu Yu, Ermo Hua, Ning Ding, Xia Hu, Bowen Zhou, Chaochao Lu, Youbang Sun

"arXiv:2608.12435v1 Announce Type: new Abstract: Transformers owe much of their strong long-context retrieval capability to a token-level memory that grows with context length. This flexibility, however, incurs a quadratic computation complexity during training and a key--value ca…"

View on X

Originally posted by Ming Zhang, Kaisen Yang, Shu Yu, Ermo Hua, Ning Ding, Xia Hu, Bowen Zhou, Chaochao Lu, Youbang Sun on X · view source

Want to go deeper?

Turn these trends into skills with Learnijoy's hands-on AI & tech courses.

Explore courses

More in AI Engineering & DevTools