A Study on the Continuous Generation of Test Cases for Large Language Models with Project Memory

<title>Abstract</title> <p>Changes in requirements, interface updates, and recurring defects make it difficult to generate continuous testing in large scale software systems. Conventional automatic scripts and large language models with no memory often fail to reuse historic failures and cover new risks. In order to improve the continuity and high risk of route coverage, we present an LLM generation approach with project memory. Based on the requirement document, the interface specification, the code commit, the defect report, and the existing script, the project memory repository is constructed, and Llama3.1-70B-Instruct is guided in the generation of executable test cases through the combination of the release awareness, the defect feedback weighting, the risk re-ordering, the structural prompts, and the semantic deduplication verification. The results of the re-release evaluation of the actual e-Commerce Order System show that the results obtained are 88.7%, the duplication reduction is 8.9%, the high risk route coverage is 91.3%, and the defect discovery is 19.6%. The results show that project memory enhances the model's understanding of long term business rules, cross-interface dependencies, and historic failure patterns.</p>