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

• URL: http://arxiv.org/abs/2603.23516v1
• Date: Fri, 06 Mar 2026 02:29:54 GMT
• Title: MSA: Memory Sparse Attention for Efficient End-to-End Memory Model Scaling to 100M Tokens
• Abstract summary: We present Memory Sparse Attention (MSA), an end-to-end trainable, efficient, and massively scalable memory model framework.<n>MSA achieves linear complexity in both training and inference while maintaining exceptional stability.<n>We also propose Memory Interleaving to facilitate complex multi-hop reasoning across scattered memory segments.
• Score: 38.762431789628074
• License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
• 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.

Related papers

• MemoryWAM: Efficient World Action Modeling with Persistent Memory [80.90899269062128]
  We introduce MemoryWAM, a world action model with efficient persistent memory.<n>A tailored attention mechanism enables retrieval of both detailed short-term context and compressed long-term context.<n>MemoryWAM outperforms strong vision-language-action (VLA) and WAM baselines in both simulation and the real world.
  arXiv  Detail & Related papers  (2026-06-18T17:59:51Z)
• LMEB: Long-horizon Memory Embedding Benchmark [49.57481835614834]
  We introduce the Long-horizon Memory Embedding Benchmark (LMEB), a comprehensive framework that evaluates embedding models' capabilities.<n>LMEB spans 22 datasets and 193 zero-shot retrieval tasks across 4 memory types.<n>We evaluate 15 widely used embedding models, ranging from hundreds of millions to ten billion parameters.
  arXiv  Detail & Related papers  (2026-03-13T02:09:57Z)
• AllMem: A Memory-centric Recipe for Efficient Long-context Modeling [32.025154452526856]
  Large Language Models (LLMs) encounter significant performance bottlenecks in long-sequence tasks.<n>We introduce textscAllMem, a novel and efficient hybrid architecture that integrates Sliding Window Attention (SWA) with non-linear Test-Time Training (TTT) memory networks.
  arXiv  Detail & Related papers  (2026-02-14T09:04:28Z)
• MemoryRewardBench: Benchmarking Reward Models for Long-Term Memory Management in Large Language Models [40.965722377085456]
  We introduce MemoryRewardBench, the first benchmark to systematically study the ability of reward models to evaluate memory quality.<n> Evaluations on 13 cutting-edge RMs indicate a diminishing performance gap between open-source and proprietary models.
  arXiv  Detail & Related papers  (2026-01-17T09:04:53Z)
• Agentic Memory: Learning Unified Long-Term and Short-Term Memory Management for Large Language Model Agents [57.38404718635204]
  Large language model (LLM) agents face fundamental limitations in long-horizon reasoning due to finite context windows.<n>Existing methods typically handle long-term memory (LTM) and short-term memory (STM) as separate components.<n>We propose Agentic Memory (AgeMem), a unified framework that integrates LTM and STM management directly into the agent's policy.
  arXiv  Detail & Related papers  (2026-01-05T08:24:16Z)
• MoM: Linear Sequence Modeling with Mixture-of-Memories [11.7218875689111]
  We introduce a novel architecture called Mixture-of-Memories (MoM)<n>MoM utilizes multiple independent memory states, with a router network directing input tokens to specific memory states.<n>MoM performs exceptionally well on recall-intensive tasks, surpassing existing linear sequence modeling techniques.
  arXiv  Detail & Related papers  (2025-02-19T12:53:55Z)
• vTensor: Flexible Virtual Tensor Management for Efficient LLM Serving [53.972175896814505]
  Large Language Models (LLMs) are widely used across various domains, processing millions of daily requests. Large Language Models (LLMs) are widely used across various domains, processing millions of daily requests.
  arXiv  Detail & Related papers  (2024-07-22T14:37:58Z)
• MEMO: Fine-grained Tensor Management For Ultra-long Context LLM Training [24.066283519769968]
  Large Language Models (LLMs) have been trained using extended context lengths to foster more creative applications.<n>We propose MEMO, a novel framework for fine-grained activation memory management.<n>MeMO achieves an average of 1.97x and 1.80x MFU compared to Megatron-LM and DeepSpeed.
  arXiv  Detail & Related papers  (2024-07-16T18:59:49Z)
• B'MOJO: Hybrid State Space Realizations of Foundation Models with Eidetic and Fading Memory [91.81390121042192]
  We develop a class of models called B'MOJO to seamlessly combine eidetic and fading memory within an composable module. B'MOJO's ability to modulate eidetic and fading memory results in better inference on longer sequences tested up to 32K tokens.
  arXiv  Detail & Related papers  (2024-07-08T18:41:01Z)
• Memformer: A Memory-Augmented Transformer for Sequence Modeling [55.780849185884996]
  We present Memformer, an efficient neural network for sequence modeling. Our model achieves linear time complexity and constant memory space complexity when processing long sequences.
  arXiv  Detail & Related papers  (2020-10-14T09:03:36Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.