Developer discovers AI agents wrote notes for months that nothing ever read

A software developer found that worker agents in an orchestration platform had been writing detailed notes files for months, but the files were never injected into any prompt or read at any stage of execution, making them effectively write-only memory. A 30-second codebase search revealed only two references to the filename — an archiving routine and a UI file-lister — confirming no system ever consumed the stored knowledge. The same platform also discarded each job's plan after every run, forcing agents to rebuild logic from scratch each time, which inflated average agent turns from an ideal five to 6.5, with worst cases reaching ten. The developer has since implemented a curated, size-bounded notes structure where a trimmed header travels with each run while the full file stays on disk, and plan instructions are always preserved over context when limits are hit. The post urges other developers to grep their codebases for filenames their agents write to, arguing this pattern of unread agent memory is a common, easy-to-miss class of bug.