Amazon's EMMA treats multi-agent configs as episodic memory, beating baselines by up to 32.5%

omarsar0 · x · 2026-10-28

Amazon and UCSD researchers published EMMA (Structured Episodic Memory for Multi-Agent Workflow Synthesis), which treats a multi-agent system's configuration as episodic memory that improves from failures.
Architecture: one LLM runs everything — a controller plans tasks step by step and picks sub-agents and skills; sub-agents start as generalists, specialists inherit the generalist prompt plus role-specific methods; skills are short procedures like algorithms and pitfall checklists.
Core mechanism: the controller's routing memory, agent prompts, and skills are all episodic memory. After a failed task, a reflector assigns ownership of the fix and proposes exactly one edit there; guards keep every accepted edit scoped, incremental, and bounded — avoiding "ill-scoped evolution".
Results: across KramaBench, AppWorld, and BBEH with four baselines (ACE, GEPA, EvoMAS, ADAS) and multiple backbones, EMMA achieves the best overall performance on every benchmark, up to 32.5% over the strongest baseline, using 17–38% fewer training executions.

More from coding & agent