Explicit, scoped episodic memory enhances multi-agent LLM system adaptation

EMMA assigns each failure to a specific system component and applies guarded single edits, consistently outperforming four baselines across three benchmarks with five language model backbones.

Big Tech

Ruijia Niu · Kaiwen Liu · Rose Yu · Kristina Klinkner

Amazon · University of California San Diego

Research Digest·28 October 2026·3 min read

Illustration: Explicit, scoped episodic memory enhances multi-agent LLM system adaptation

In one line

EMMA treats multi-agent configurations as episodic memory with explicit, guarded updates, outperforming baselines by up to 32.5%.

Before you read this

Three ideas this paper builds on

agent memoryA repository where an agent stores past interactions and outcomes to inform future decisions.196 papers →multi-agent architectureA system where multiple AI agents collaborate, each handling specialized roles, communicating and coordinating to complete complex tasks.96 papers →failure attributionIdentifying which component or step in a system caused an observed failure, often using structured trace representations.111 papers →

Read it as

Deploying AI agents for specialized tasks like data analysis often needs knowledge that general models lack. Training is costly and opaque; hand-designing agents demands expert labor.

EMMA treats the configuration of a multi-agent system (which agent decides the next step, what skills each agent has) as a kind of memory. When a failure is diagnosed, it makes a single, small change in only one part of the system, and that change must pass rules that keep it focused and incremental.

Across three benchmarks and five different models, EMMA outperformed the strongest competing method by up to 32.5% and used far fewer training runs. This shows that agents can learn from their mistakes more effectively when updates are scoped and checked, rather than relying on the model's own preferences.

How it works

stage by stage

RSS

The Daily Edition

The front page in your inbox each weekday morning: the lead, the picks, and why each one earned its place.

Concepts

agent memory multi-agent architecture failure attribution agent skills model routing multi-agent coordination