Why does GPT sometimes lose track of requirements in a long conversation?

I use GPT for work that involves several constraints, such as keeping a particular format while following project-specific terminology. In a long chat, it may handle the latest request well but miss something established earlier.

Is this mainly a context-length issue, or can the way requirements are phrased and updated make a difference? I’d also be interested in how people test whether a model has retained the important constraints before relying on its output.