Where does the goal lose its memory?

The bundled CSV has eight rows grouped into three workflows. Lead qualification has three steps, content sprint three and follow up two. Within each workflow adjacent handoffs count one fewer than rows: two, two and one, for five total. At forty pixels per count rows, workflows and handoffs measure three hundred twenty, one hundred twenty and two hundred. Numeric steps order the rows and transitions never cross workflow boundaries. These are supplied labels and durations, not observed remote agent behavior or a verified productivity study. Each handoff compares agent, workspace and memory labels. Four sample handoffs change all three dimensions, while the follow up handoff changes none. This gives twelve changed slots among fifteen possible, leaving three unchanged. At twenty pixels per slot changed, unchanged and possible measure two hundred forty, sixty and three hundred. Dividing twelve by fifteen gives eighty percent. Four handoffs change memory. Renaming a surface label can change the score even when actual context sharing is identical, so this ratio is a property of the CSV rather than a model quality or security metric. The supplied minutes sum to thirteen for lead qualification, twenty for content sprint and seven for follow up, forty total. At eight pixels per minute the bars measure one hundred four, one hundred sixty and fifty six. Those are user supplied durations, not elapsed runtime, cost or savings. The map needs a local CSV parser and optionally a graph library. In this isolated preview their companion files are missing, so analysis fails closed rather than fabricating a result; native load sample and reset still preserve the source behavior. No provider or remote login is used and no successful export is claimed without the parser.