SUMONE: A Simple and Efficient Summarizer for Plug-and-Play Agent Memory

AI Digest

SUMONE通过自监督学习压缩代理记忆，减少推理成本，提升效率。

SUMONE introduces a self-supervised summarizer for agent memory compression, reducing inference costs by 78% in DeepSearchQA and 75% in SWE-bench Verified without task-specific fine-tuning.

Key points

• SUMONE通过掩码重建实现任务无关的自监督训练 SUMONE uses masked reconstruction for task-agnostic self-supervision
• 压缩后推理成本降低78%-75%，效果优于全上下文 Achieves 78-75% cost reduction vs full context inference
• 无需下游任务微调即可跨模型迁移应用 Transfers across models without additional fine-tuning
• 保留轻量级摘要器实现插件式上下文压缩 Retains lightweight summarizer for plug-and-play compression

Takeaway: SUMONE为代理记忆压缩提供高效自监督方案，显著降低计算成本。 / SUMONE offers efficient self-supervised memory compression for agents, substantially reducing computational costs.

Why it matters 为长时序代理提供高效压缩方案，降低计算成本，提升实际应用可行性。