Coding Agents Never Voluntarily Read Their Own Memory: 0 Memory Operations in 114 Turns Even With a Pre-Seeded Store

arXiv 2607.20972·high signal

Swapnanil Saha argues agent memory must be a harness property delivered involuntarily, not a document the agent chooses to read, and backs it with a controlled evaluation on a real coding task: an agent with a pre-seeded memory store performed 0 memory operations across 114 turns, while deterministic cue-triggered injection delivered in every seeded run with zero false alarms. A repeated-compaction decay probe found ten facts held only in conversation vanished at the first summary and stayed absent from 106 of 108 compactions — the deprived agent resorted to grepping the harness's own session files to rebuild them — while harness-injected facts survived all 138 compact-resumes. The paper also measures that 39% of intra-session re-reads re-buy content already paid for before a compaction boundary.