MemoSync: Co-Managing Developer-Agent Metamemory Across Contexts
Baiqiao Zhang, Junze Li, Chunyang Luo, Yuying Tang, Wenshuo Zhang, Shiyuan Piao, Dingdong Liu, Shuai Ma, Xiaojuan Ma
Show agent memory as a reviewable list the developer can edit and scope. Include a record of which items the agent actually used. Developers did more monitoring without extra workload.
AI coding agents carry context across sessions, but developers cannot see what gets stored or used. Their judgments about memory drift from the agent's actual memory.
Method: MemoSync makes agent memory a shared, editable list. The agent proposes memory items and reports which ones it used. The developer reviews, edits, and limits their scope. In a between-subjects study with 132 participants against two baselines, MemoSync users had significantly higher task performance and the highest developer-agent memory alignment. A mediation analysis tied the gain mainly to better monitoring. Users did not report higher workload despite many more actions.
Caveats: The mediation analysis was exploratory. The week-long deployment included only nine developers.
Reflections: Does the benefit hold when memory lists grow to hundreds of items across many projects? · Do developers keep monitoring memory after the novelty of a shared list wears off?