Pick a question. The same AI answers it twice: once pointed at the company’s files, and once connected to its memory.

Holcombe Mechanical is a fictional $16M commercial mechanical contractor. On the left, the AI searches Holcombe’s shared drive, inboxes, and system exports as they stand today: 8,700 files going back to 2020, 13 MB of text, with the superseded vendor quote filed beside the current one, four copies of the estimating workbook, and the warranty position spread across three documents. It can search the drive and open its files, at most 25 times for each question, and it has the plain request to answer from the files. On the right, the same AI reads the company memory: 27 entries, each approved by a named person or kept with its source, and connected to the customers, jobs, and rules it concerns. It gives the position somebody decided, with the record behind it. Each answer was graded against a key written from Holcombe’s records: the files version was right on 12 of 20 questions, and the memory on 20 of 20.

The answers are real model output, produced by Claude Code agents on 2026-09-09 (model claude-opus-5): for each question one agent searched the drive and one read the memory, and each wrote its answer to a file that is replayed here without edits. Each answer was then checked for the facts in its key and judged against the key by the same model; the verdict sits under the answer. No model runs in your browser.