# A Human-Inspired Reading Agent with Gist Memory of Very Long Contexts

**Authors:** Kuang-Huei Lee, Xinyun Chen, Hiroki Furuta, John Canny, Ian Fischer  
**arXiv:** [2402.09727](https://arxiv.org/abs/2402.09727)  
**Date:** 2024-02-15  
**Summary page:** https://www.summarizepaper.com/en/paper/2402.09727v1/

## Key points

- The paper addresses limitations of current Large Language Models (LLMs) in handling long inputs.
- ReadAgent is proposed as an LLM agent system that increases the effective context length up to 20 times in experiments.
- ReadAgent operates in three primary steps: episode pagination, memory gisting, and interactive look-up.
- Evaluation against baselines shows that ReadAgent outperforms all across challenging long-document comprehension tasks.
- ReadAgent can be adapted for web navigation settings with very-long contexts, showing promising performance results.
- Primary contributions include introducing ReadAgent as a human-inspired LLM agent, demonstrating significant performance advantages through experimental evaluations, comparing against popular baselines, and providing detailed analysis of results.

## Summary

The paper "A Human-Inspired Reading Agent with Gist Memory of Very Long Contexts" by Kuang-Huei Lee, Xinyun Chen, Hiroki Furuta, John Canny, and Ian Fischer from Google DeepMind and Google Research addresses the limitations of current Large Language Models (LLMs) in handling long inputs. The authors propose ReadAgent, an LLM agent system that significantly increases the effective context length up to 20 times in experiments. <br>
Inspired by how humans interactively read long documents, ReadAgent is designed as a prompting system that leverages the advanced language capabilities of LLMs. The system operates in three primary steps: episode pagination where the LLM decides where to pause in reading text to create episodes or pages; memory gisting where each page is compressed into a shorter gist memory associated with its context; and interactive look-up where the LLM retrieves relevant information from raw text to solve tasks. <br>
The authors evaluate ReadAgent against baselines using retrieval methods, original long contexts, and gist memories on challenging long-document comprehension tasks such as QuALITY, NarrativeQA, and QMSum. ReadAgent outperforms all baselines across these tasks while extending the effective context window by 3-20 times. Additionally, the paper demonstrates how ReadAgent can be adapted for web navigation settings with fundamentally very-long contexts. The authors find promising performance results in this setting as well. <br>
Overall, the primary contributions of this work are introducing ReadAgent as a human-inspired LLM agent that generates gist memories and looks up information as needed for solving tasks on long contexts; demonstrating significant performance advantages and scalability through experimental evaluations on challenging benchmarks; comparing against popular baselines; and providing detailed analysis of results.