Tag
This paper shows that for coding agents' memory systems, checking whether a specific claim remains valid after a repository change provides higher precision than assessing behavior preservation in diffs, validated through experiments with multiple LLMs and real-world data.