MSA: Memory Sparse Attention for Efficient End-to-End Memory Model Scaling to 100M Tokens

2026-03-06

AI Summary

Current large language models struggle to efficiently process ultra-long contexts due to their full attention mechanisms, suffering from degraded accuracy, high latency, non-dynamic memory updates, and a lack of end-to-end optimization. This work proposes the Memory Sparse Attention (MSA) framework, which integrates scalable sparse attention, document-level RoPE, KV cache compression, memory parallelism, and memory interleaving to achieve, for the first time, an end-to-end trainable memory-augmented model capable of dynamic memory modification and multi-hop reasoning over 100 million tokens. The approach incurs less than 9% performance degradation during inference on just two A800 GPUs and significantly outperforms existing large language models, retrieval-augmented generation (RAG) systems, and memory-based agents on long-context benchmarks.

Technology Category

• Machine Learning: Large Multimodal Models (LMMs)
• Natural Language Processing: (Large) Language Models
• Computer Vision: Large Vision Models

Application Category

• Search and Retrieval-Augmented AI: Large language models for search
• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs
• User Modeling, Personalization and Recommendation: Large Language Models (LLM) for user modeling and recommendation

Abstract

Long-term memory is a cornerstone of human intelligence. Enabling AI to process lifetime-scale information remains a long-standing pursuit in the field. Due to the constraints of full-attention architectures, the effective context length of large language models (LLMs) is typically limited to 1M tokens. Existing approaches, such as hybrid linear attention, fixed-size memory states (e.g., RNNs), and external storage methods like RAG or agent systems, attempt to extend this limit. However, they often suffer from severe precision degradation and rapidly increasing latency as context length grows, an inability to dynamically modify memory content, or a lack of end-to-end optimization. These bottlenecks impede complex scenarios like large-corpus summarization, Digital Twins, and long-history agent reasoning, while limiting memory capacity and slowing inference. We present Memory Sparse Attention (MSA), an end-to-end trainable, efficient, and massively scalable memory model framework. Through core innovations including scalable sparse attention and document-wise RoPE, MSA achieves linear complexity in both training and inference while maintaining exceptional stability, exhibiting less than 9% degradation when scaling from 16K to 100M tokens. Furthermore, KV cache compression, combined with Memory Parallel, enables 100M-token inference on 2xA800 GPUs. We also propose Memory Interleaving to facilitate complex multi-hop reasoning across scattered memory segments. MSA significantly surpasses frontier LLMs, state-of-the-art RAG systems, and leading memory agents in long-context benchmarks. These results demonstrate that by decoupling memory capacity from reasoning, MSA provides a scalable foundation to endow general-purpose models with intrinsic, lifetime-scale memory.

Problem

• long-term memory
• context length
• memory scaling
• attention mechanism
• end-to-end optimization

Innovation

• Memory Sparse Attention
• long-context modeling
• scalable sparse attention
• document-wise RoPE
• Memory Interleaving