AI agents that write code for hours at a stretch have a memory problem: they cannot always tell which of their own earlier notes a later edit has already made false.
Researchers propose StateTape, a framework that rewrites a coding agent's context as the repository changes instead of just trimming it as it grows. Instead of guessing staleness from the text of old observations, StateTape builds a symbol-level code graph of the repo, using dependencies and language rules to know exactly which symbols a given write can affect. A tape then marks which symbols each write touched, and a per-write process uses that tape, plus a small manager model for judgment calls, to clear out records a write has falsified and keep the ones that still hold. The team also built TraceBench, a benchmark that labels what an agent is holding in context against what it actually needs next, to measure this directly.
The headline metric is "resolve rate": how often an agent, at a given step, is holding the code information it actually needs rather than outdated or irrelevant leftovers. Tested across six different coding agents and three edit-heavy benchmarks, StateTape improved resolve rate in every combination, with what the authors describe as little added computational overhead. The paper's abstract does not spell out the exact percentage gains, so the size of the improvement over plain history-pruning is still an open question.
Current context-management tricks - summarizing or pruning old text - treat code like prose, blind to the fact that one edit can quietly break the meaning of code written five files and two hours earlier. Tying memory management to an actual dependency graph is a more principled fix, but principled and proven-at-scale are different things, and three benchmarks is a modest sample size to hang that claim on.