Track trials, findings, and clinical evidence as it lands.
Search papers, labs, and topics across Lattice.
This study investigates the heterogeneous nature of working memory in coding agents, emphasizing how different semantic roles of memory objects influence their retention and compression behaviors. By analyzing 55 coding-agent trajectories, the authors reveal that semantically distinct objects require tailored memory management strategies, such as an object-aware compression policy and a retrieval-based policy. The findings indicate that traditional token budgets may misrepresent the true management costs and context delivery, highlighting the need for a nuanced approach to evaluating memory-management strategies.
Semantically distinct memory objects in coding agents reveal that equal token budgets can mislead about context delivery and management costs.
Agent working memory is heterogeneous. Objects such as instructions, artifacts, tool outputs, and agent-generated state play different semantic roles and exhibit different size, retention, and representation profiles. Recent work has begun to explore memory-management mechanisms that account for such heterogeneity. This work focuses on semantic heterogeneity and studies how it should shape the management and evaluation of working memory in coding agents. Across 55 archived coding-agent trajectories, we find that semantically different working-memory objects exhibit distinct retention and compression behavior. This heterogeneity motivates semantically informed memory management. We study two semantically informed strategies: an object-aware compression policy and a retrieval-based policy. Their evaluation shows 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. Together, these results show why semantic structure matters for agent working memory and why evaluating memory-management strategies requires more than a nominal token budget. We organize these lessons into four levels: stored state, delivered context, management work, and task or process outcome.