The people outside the memory

The owner looked at a new REM notebook and asked where everyone was. The first five-day trial had been narrowed further for a page test: --first-people 0 meant it wrote no person memories at all. Removing that cap would still not answer the larger question. The normal map read only ninety days of mail, and one short exchange did not earn a Markdown page. Those contacts existed in private metadata but disappeared from the People view.

Making a page for every address would repeat an earlier failure. A real notebook once held hundreds of empty person templates, including newsletters and the owner's other inboxes. A mail header is evidence that an address was seen; it is not proof of who a person is, or what the relationship means. The new directory keeps those claims apart. A person REM has written about remains in the main table. A one-off correspondent can be found in a separate, searchable list, marked as an unreviewed contact, with a terminal command to prepare a memory when the owner recognizes them.

The history question has a cost before a model writes anything. An isolated metadata-only scan of one connected Outlook account traversed its available years since 1970 in 461 seconds, but Gmail timed out before its address list was read. REM called that first result a lower bound. After the Gmail path switched to bounded direct metadata reads, the two-mailbox scan finished in 844 seconds. It observed 22,411 headers and mapped 634 People records and ten projects. Another 803 low-signal contacts stayed searchable for review. Both providers reported zero incomplete windows. The directory still explains coverage because a later mailbox or time window can fail. A timed-out window is split, and any day still unavailable is recorded rather than counted as scanned.

The combined map offers 589 person investigations, eight projects and the owner's page. The current defaults estimate 660 million billed input tokens and about 297 minutes for those 598 pages with ten workers. Adding mapped organizations and installed Skills brings the selected queue to 886 pages, 751 million input tokens and 364 minutes. There are no completed page samples in this isolated notebook, so those are planning numbers, not promises or weekly quota percentages. The configured investigation target is advisory; the weekly safety floor still stops new model work. REM can show the forecast without a model turn or a historical body archive. After a foreground init, it also offers the source and schedule approval needed for nightly upkeep without repeating the first batch. A full all-person run and claim-by-claim review remain separate work, because finding an address is only the start of remembering a person well.