# EMMA's scoped memory edits beat multi-agent baselines on three benchmarks

> Researchers from UC San Diego and Amazon report EMMA has the best overall performance on KramaBench, AppWorld and BBEH across five backbones.

- Status: official (From the company, authors or authority themselves, with a link to their own page.)
- Topic: Research
- Published: 2026-11-01T17:34:56+00:00
- Updated: 2026-11-01T17:34:56+00:00
- Web page: https://ai-news.ikangai.com/story/emma-s-scoped-memory-edits-beat-multi-agent-baselines-on-three-benchmarks/
- Seen in: DAIR.AI (Elvis Saravia)
- Change: research (Research)

Researchers from UC San Diego and Amazon report that EMMA, a method for improving multi-agent systems, posts the best overall performance on KramaBench, AppWorld and BBEH. It was compared against four baselines: ACE, GEPA, EvoMAS and ADAS. The margin reaches 32.5% over the strongest baseline. EMMA ranks first in 17 of 20 settings across five backbones, while using 17 to 38% fewer training executions than the most efficient baseline.

The core idea is to treat the configuration of a multi-agent system as episodic memory. In EMMA's design, a controller plans each task one step at a time and draws on routing memory. Specialist agents inherit a generalist prompt, and skills hold short per-step procedures. All three levels are text run by one fixed LLM, so the configuration changes as text while the same model runs it. The authors argue that earlier multi-agent evolution methods let the backbone model decide where and how much the system changes. As a result, learning follows the model's habits: one backbone accumulates skills while another spawns agents.

EMMA assigns ownership of each change. After a failure, a reflector attributes the diagnosed failure to the scope that owns its correction and proposes one edit within that scope. Guards that do not depend on the backbone then keep each accepted edit scoped, incremental and bounded. Because these guards are backbone-independent, they do not depend on which model runs underneath.

Multi-agent search baselines fall sharply on open-weight models. On the AppWorld challenge, EvoMAS scores 89.0 with Opus 5 and 7.2 with GLM-5. EMMA's weakest backbone still reaches 61.2.

The authors note a caveat: accepted edits persist even when a task stays unsolved, so ineffective entries may remain in memory.

## Why it matters

Baselines that perform well on frontier models can fall sharply on open-weight models, so agent-tuning methods need testing across backbones. The authors note that accepted edits persist even when a task stays unsolved.

## What you can do

Before adopting an automated agent-tuning method, test it on the open-weight models you would actually deploy.

## Original source

- [(2610.35966) EMMA: Structured Episodic Memory for Multi-Agent Workflow Synthesis](https://arxiv.org/abs/2610.35966)
- [EMMA: Structured Episodic Memory for Multi-Agent Workflow Synthesis | DAIR.AI Academy | DAIR.AI Academy](https://academy.dair.ai/papers/emma-structured-episodic-memory-for-multi-agent-workflow-synthesis-2610.35966)

---

AI wrote this summary from the sources listed, and nobody reviewed it individually before publication. Check the original source before you rely on it.

Cite as: IKANGAI AI News, "EMMA's scoped memory edits beat multi-agent baselines on three benchmarks", 2026-11-01, https://ai-news.ikangai.com/story/emma-s-scoped-memory-edits-beat-multi-agent-baselines-on-three-benchmarks/
