Equal Token Budgets Do Not Mean Equal Delivered Context: 55 Coding-Agent Trajectories Say Memory Is Semantically Heterogeneous

arXiv 2608.31057·medium signalXBlueskyLinkedInCopy link

Analyzing 55 archived coding-agent trajectories shows instructions, artifacts, tool outputs and agent-generated state have distinct size, retention and compression profiles, so treating working memory as one undifferentiated token budget misprices it. Testing an object-aware compression policy and a retrieval-based policy, calibration gains do not transfer to held-out tasks, and equal nominal token budgets produce unequal delivered context and unequal management cost, with a real-system replay exposing serving limits the budget alone hides. The paper proposes evaluating memory management at four levels, stored state, delivered context, management work, and task outcome, instead of quoting a context-window number.