I use the Boy Scout rule to pick a baseline error rate for new releases: leave each deployment at least as clean as you found it ✌️ My error rate might not be zero, but it shouldn't creep up between deployments.

Agent memory is only useful for interactive sessions. Background agents need reproducibility, and memory breaks that. It also confounds your evals by patching in context just-in-time. If a background agent would benefit from memory, that's a signal to refresh your prompt instead.