EvoGroup is a system designed to help AI agents retain knowledge shared by a team, such as group relationships and information built collaboratively. Instead of relying on a fixed, hand-written memory setup, it starts with six tools and revises its approach based on past task outcomes and uncertain cases.
In the researchers’ EverMemBench test, pass@1 rose from 71.07% to 89.58%, above the 85.83% reported for Claude Code. They also say a version of the harness frozen after development performed well on the separate GroupMemBench, using GPT-5.5; the abstract gives no score for that test. The researchers say they plan to open-source the code.
Why it matters
For teams considering AI agents that work across shared discussions, the results suggest an adaptive memory setup may help preserve group knowledge, but the abstract does not provide GroupMemBench scores or independent validation.
