# Unleashing Infinite-Length Input Capacity for Large-scale Language Models with Self-Controlled Memory System

**Authors:** Xinnian Liang, Bing Wang, Hui Huang, Shuangzhi Wu, Peihao Wu, Lu Lu, Zejun Ma, Zhoujun Li  
**arXiv:** [2304.13343](https://arxiv.org/abs/2304.13343)  
**Date:** 2023-04-26  
**Summary page:** https://www.summarizepaper.com/en/paper/2304.13343v1/

## Key points

- The authors propose the Self-Controlled Memory (SCM) system to address the limitation of Large-scale Language Models (LLMs) in processing lengthy inputs.
- The SCM system is composed of three key modules: the language model agent, the memory stream, and the memory controller.
- The language model agent iteratively processes ultra-long inputs and stores all historical information in the memory stream.
- The memory controller provides both long-term memory and short-term memory to generate precise and coherent responses.
- SCM system can be integrated with any LLMs to enable them to process ultra-long texts without any modification or fine-tuning.
- Experimental results show that their SCM system enables LLMs to achieve multi-turn dialogue capabilities comparable to ChatGPT and outperform ChatGPT in scenarios involving ultra-long document summarization or long-term conversations.
- Limitations in evaluating the handling of extremely lengthy texts due to a lack of appropriate datasets for comprehensive and objective evaluation.
- Future work will focus on releasing a comprehensive test set and its manual evaluation criteria while testing their system on various open-source models currently available.

## Summary

In this paper, the authors propose the Self-Controlled Memory (SCM) system to address the limitation of Large-scale Language Models (LLMs) in processing lengthy inputs. The SCM system is composed of three key modules: the language model agent, the memory stream, and the memory controller. The language model agent iteratively processes ultra-long inputs and stores all historical information in the memory stream. The memory controller provides both long-term memory and short-term memory to generate precise and coherent responses. The authors demonstrate that their SCM system can be integrated with any LLMs to enable them to process ultra-long texts without any modification or fine-tuning. Experimental results show that their SCM system enables LLMs to achieve multi-turn dialogue capabilities comparable to ChatGPT and outperform ChatGPT in scenarios involving ultra-long document summarization or long-term conversations. However, there are limitations in evaluating the handling of extremely lengthy texts due to a lack of appropriate datasets for comprehensive and objective evaluation. Therefore, the authors aim to construct a specific test set that incorporates various key indicators essential for processing long texts in diverse settings. Additionally, they plan to assess the efficacy of their system on more open-source models that possess single-turn instruction comprehension capability. In conclusion, this paper proposes an effective method for extending input length for LLMs without requiring any training or modification of models. Future work will focus on releasing a comprehensive test set and its manual evaluation criteria while testing their system on various open-source models currently available.