Recent coding agents built on large language models have achieved impressive results, yet they struggle with long‑horizon tasks because each observation is appended to the context, causing unbounded growth. A common remedy is history‑based maintenance, which masks, summarizes, or prunes old observations to keep the context bounded. However, this approach relies solely on the raw text and ignores the structural connections among code symbols. Since a coding agent may edit the same codebase many times, any single write can change the meaning of symbols elsewhere; traditional maintenance may retain records falsified by a write, discard still‑valid ones, and miss code the agent will need next.
To address these issues, the paper introduces StateTape, a scalable framework that rewrites an agent’s context as the repository evolves rather than letting it grow indefinitely. The key idea is to model the repository as a symbol‑level code graph; the graph’s dependencies and language rules expose which symbols a write can affect. On top of this graph, a tape marks the symbols changed by each write, turning staleness from an inference about text into a direct observation of the agent’s writes.
StateTape’s per‑write procedure works as follows: (1) the tape, guided by the code graph, nominates records that the current write could have falsified; (2) a lightweight manager model resolves the ambiguity for records the tape cannot decide. The authors provide a theoretical analysis showing that, under reasonable assumptions, the mechanism reliably clears invalidated records and restores needed evidence.
For evaluation, the authors construct TraceBench, a benchmark that labels what evidence an agent actually holds against what is truly required for the task. Experiments span six coding agents and three edit‑heavy benchmarks; results demonstrate that StateTape effectively removes falsified records, retrieves necessary information, achieves higher resolve rates across all settings, and incurs minimal computational overhead.
Review