Self evolving group agents improve shared memory handling
EvoGroup: A Self-Evolving Harness for Group Agents in Shared Contexts
Computation and Language
Summary
Group agents help multiple people work together by keeping track of shared information. However, current systems struggle to remember group knowledge and relationships well. The authors designed EvoGroup, a system that learns and improves over time by picking useful information, reflecting on uncertainties, and refining itself. Their tests show EvoGroup remembers group information better than other methods and works well even with new groups and tasks.
What this means in practice
- •For collaboration platform developers: Improve group agent memory to better capture team knowledge and relationships in shared workspaces.
- •For enterprise software teams: Enhance digital assistants to adaptively evolve with ongoing multi-user interactions for long-term projects.
Authors
Xinyi Xu, Xiaoyang Li, Yucheng Feng, Wenqing Chen, Youliang Yuan, Wei Shi, Rongxiang Weng, Zibin Zheng
Abstract
With the rise of Group Agent such as Claude Tag, agents increasingly serve as digital workers for long-term team collaboration. These agents operate over shared contexts that continuously accumulate through multi-user interactions. However, existing agent memory systems are mainly designed for individual users and struggle to capture group-level knowledge in shared contexts, including group relationships and collaboratively formed knowledge. Manual memory strategies further limit adaptation to diverse and evolving shared contexts. To this end, we propose \textbf{EvoGroup (EvoG)}, a self-evolving harness for group agents that starts from a minimal six-tool cold start and adapts through interaction with shared context. EvoG uses confidence and task outcomes to select informative trajectories, while also eliciting self-reflections on uncertainty in low-confidence cases. The analysis agent then examines selected trajectories through adaptive progressive disclosure, followed by bucketed analysis to synthesize high-level findings that guide the evolve agent in revising the Harness. Empirically, pass@1 on held-out split of EverMemBench improves from 71.07\% to 89.58\%, surpassing both self-evolving memory and human-designed harnesses, including Claude Code (85.83\%). To evaluate generalization, we freeze the evolved harness and evaluate it on the unseen GroupMemBench, where, with GPT-5.5, it demonstrates effective generalization and outperforms both memory baselines and human-designed harnesses.