Auto-memory audit of 41 files finds MEMORY.md orphan

The founder's problem started with a record created at a time when it was true: an X ad campaign that appeared to spend ¥6,000 with zero sales. Later, checking Amazon Associate tracking, he found 4 orders and ¥7,071 revenue, an ROI +18% — meaning the original memory was a measurement gap, not a failure. If that single file stayed, the AI would never again propose X ads.

The deeper issue is that Auto-memory files carry a description used for recall, and vague descriptions cause memories to be pulled into unrelated situations. The founder's approach is to treat memory as inventory rather than an asset — something to be periodically counted, not trusted indefinitely. He uses grep to surface files containing dates and amounts, and an AI subagent to find contradictions, duplicates, and unsupported claims, but reserves deletion for himself.

What this hinges on is whether detection and correction stay separate. He once let an AI both find and fix issues and a correct memory was deleted as a duplicate; recovery was lucky, since memory directories are normally outside git. Whether other solo operators adopt a similar routine may depend on how costly one wrong memory becomes in their own workflow.