EMMA: Amazon's self-improving multi-agent system using episodic memory

Amazon and colleagues publish EMMA, a self-improving multi-agent system in which one LLM runs a controller, sub-agents, and skills that all update from failures.

After a failed task, a reflector decides which component owns the fix and proposes one bounded edit. Elvis Saravia reports that EMMA beats ACE, GEPA, EvoMAS, and ADAS on KramaBench, AppWorld, and BBEH, by up to 32.5% over the strongest baseline, while using 17 to 38% fewer training executions. The figures come from the paper as summarized in Saravia's post; the report does not include independent verification.