ReadAgent is an LLM agent system that improves the effective context length for long-document reading comprehension. Inspired by human reading behavior, ReadAgent uses a prompting system to store, compress, and retrieve information from long texts. It divides long texts into episodes, compresses each into a gist memory, and allows the LLM to look up relevant passages when needed. ReadAgent outperforms baselines on three long-document reading tasks—QuALITY, NarrativeQA, and QMSum—while significantly increasing the effective context length. It achieves a 3.5–20× increase in effective context length, with notable improvements in performance metrics such as LLM rating and ROUGE-L. ReadAgent is simple to implement and can be built using existing LLMs without requiring architectural changes or additional training. It is also adaptable to web navigation tasks. The system uses a combination of episode pagination, memory gisting, and interactive look-up to efficiently process long texts. ReadAgent demonstrates strong performance in both long-context reading comprehension and tasks requiring retrieval of specific information from long documents. The system's effectiveness is supported by extensive experimental results across multiple benchmarks and tasks.ReadAgent is an LLM agent system that improves the effective context length for long-document reading comprehension. Inspired by human reading behavior, ReadAgent uses a prompting system to store, compress, and retrieve information from long texts. It divides long texts into episodes, compresses each into a gist memory, and allows the LLM to look up relevant passages when needed. ReadAgent outperforms baselines on three long-document reading tasks—QuALITY, NarrativeQA, and QMSum—while significantly increasing the effective context length. It achieves a 3.5–20× increase in effective context length, with notable improvements in performance metrics such as LLM rating and ROUGE-L. ReadAgent is simple to implement and can be built using existing LLMs without requiring architectural changes or additional training. It is also adaptable to web navigation tasks. The system uses a combination of episode pagination, memory gisting, and interactive look-up to efficiently process long texts. ReadAgent demonstrates strong performance in both long-context reading comprehension and tasks requiring retrieval of specific information from long documents. The system's effectiveness is supported by extensive experimental results across multiple benchmarks and tasks.