What they did
The authors examined 55 archived coding-agent trajectories, separating working memory into semantic object types such as instructions, artifacts, tool outputs, and agent-generated state. They measured how these categories differed in size, representation, retention, and compression behavior.
They then evaluated two semantically informed management strategies: an object-aware compression policy and a retrieval-based policy. The evaluation included held-out tasks and a replay in a real serving system, and considered four levels of measurement: stored state, delivered context, management work, and task or process outcome.
Key findings
- Different semantic object types showed distinct retention and compression profiles, indicating that coding-agent memory is not a uniform pool of interchangeable tokens.
- Improvements observed while calibrating a memory policy did not necessarily transfer to held-out tasks.
- Equal nominal token budgets did not guarantee equal amounts or kinds of context delivered to the agent, nor equal computational management costs.
- Real-system replay revealed serving constraints that were not captured by nominal context or storage budgets alone.
Why it matters
The study argues that memory policies for coding agents should be evaluated by what information reaches the model, what work is required to manage it, and how the agent performs—not merely by how many tokens are stored. This framework could support more meaningful comparisons among compression, retrieval, and context-management systems.
Caveats
The empirical analysis covers only 55 archived coding-agent trajectories, and the abstract does not establish broad gains in coding success from either tested policy. Results may depend on the agents, tasks, serving stack, and semantic categories used, so larger and more diverse evaluations are needed.