Akashic: A Low-Overhead LLM Inference Service with MemAttention

2026-07-06

AI Summary

This work addresses key challenges in long-context reasoning with large language models, including high historical replay costs, limited context lengths, and dilution of critical information. To overcome these issues, the authors propose Akashic, a system that partitions context into bounded blocks and models inter-block semantic relationships via a novel MemAttention mechanism, preserving cross-block evidence without requiring full history rewriting. Akashic further integrates hardware-software co-designed memory layout optimizations to minimize retrieval fragmentation and I/O overhead. Experimental results across four representative workloads and three model scales demonstrate that Akashic achieves up to a 10.2-point improvement in task accuracy, a 1.21× increase in throughput, and a 1.88× enhancement in sustainable request rate compared to strong baselines.

Abstract

Recent LLM-based agent systems continuously accumulate context across multi-turn interactions, tool invocations, and cross-session workflows. Replaying the full history for every request quickly becomes impractical: long contexts increase prefill cost, may exceed context limits, and often bury task-relevant evidence in irrelevant content, degrading both serving efficiency and output quality. We propose Akashic, a low-overhead memory system built around MemAttention, which organizes context into bounded chunks and models semantic relationships across chunks, preserving cross-chunk evidence without repeatedly rewriting the full history. Akashic further applies hardware-software co-designed memory placement to co-locate likely co-retrieved chunks, reducing retrieval fragmentation and I/O overhead. Across four representative workloads and three model sizes, Akashic improves task accuracy by up to 10.2 points, throughput by up to 1.21x, and sustainable request rate by up to 1.88x over strong prior memory baselines.

Problem

• LLM inference
• long-context management
• memory overhead
• context accumulation
• retrieval efficiency

Innovation

• MemAttention
• context chunking
• hardware-software co-design
• memory system
• LLM inference