EVOGROUP: A SELF-EVOLVING MEMORY HARNESS FOR GROUP AGENTS IN SHARED CONTEXTS

With the rise of Group Agents 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 EvoGroup (EvoG), a self-evolving memory harness for group agents that starts from a minimal five-tool cold start and adapts through interaction experience. EvoG uses confidence and task outcomes to identify informative trajectories, while low-confidence cases provide self-reflections on uncertainty. An analysis agent then examines selected trajectories with adaptive progressive disclosure and aggregates recurring patterns into high-level findings, which guide an evolve agent to revise the Harness. Empirically, after five Harness updates, pass@1 on EverMemBench improves from 71.07% to 89.58%, surpassing both self-evolving memory methods and human-designed harnesses, including Claude Code (85.75%). For DeepSeek-V4-Flash, the evolved Harness also reduces average per-question tool calls and answer time. To test generalization, we freeze the evolved Harness and directly evaluate it on the unseen GroupMemBench, showing effective generalization while outperforming self-evolving memory baselines and Claude Code.