The paper studies working memory in , focusing on semantic heterogeneity. Objects such as instructions, artifacts, tool outputs, and agent state differ in size, retention, and compressibility. Across 55 archived trajectories, the authors find that these object types exhibit distinct behavior, motivating memory management strategies that account for semantics.
The authors evaluate two semantically informed strategies: an object-aware compression policy and a retrieval-based policy. Their results show that calibration gains may not transfer to held-out tasks, and that equal token budgets do not imply equal delivered context or management cost. A real-system replay further exposes serving limits that nominal budgets alone do not capture.
These findings matter because they demonstrate that naive token budgets are insufficient for evaluating strategies. The work offers a four-level framework (stored state, delivered context, management work, and task outcome) to guide more meaningful evaluation.