Keep the spreadsheet, not just the summary of it.

A research agent spends an hour on a question and produces two things: a clear write-up, and the artifact behind it — the spreadsheet it built, the CSV it pulled, the chart it rendered.

The write-up gets saved, because prose is easy to save. The artifact goes to a downloads folder with a name like export-3.csv, or into a chat attachment that expires, or nowhere at all because the agent had no path for it. Three weeks later the write-up says "revenue concentration is worse than we assumed" and the workbook that made that true is gone.

Now the claim isn't verifiable, and it isn't extendable either. Nobody can ask the obvious follow-up question, because answering it means redoing the analysis from the beginning.

Store the artifact where the note is

In Markbase a non-Markdown file is a first-class document at an ordinary workspace path. research/q2-model.xlsx sits directly beside research/q2-notes.md, and the note links to it with an ordinary Markdown link — no attachment model, no special syntax, nothing to remember.

Because it's a document, everything that applies to documents applies to it. It's versioned, so a model rebuilt each month keeps its earlier states. It goes in the trash and comes back out. It's covered by the same access rules and lands in the same workspace export.

Spreadsheets, PDFs, CSV extracts, rendered charts, HTML reports — the stuff agents actually produce when the answer isn't a paragraph.

Files are kept and downloaded, never rendered

Every download is served as an attachment, whatever the file claims to be. An uploaded .html or .svg is bytes you get back, not something that executes anywhere — which is a deliberate design choice rather than a limitation being explained away, and it's why storing whatever an agent produces doesn't come with a security conversation attached.

The practical consequence: no previews, no thumbnails. You download the workbook and open it in the thing that opens workbooks.

The note is what makes the file findable

Full-text search covers your Markdown. It doesn't read inside a spreadsheet — nobody's parsing your workbook's cells — so a file's discoverability comes from the note beside it: what it is, what question it answers, what the method was, what to be careful about.

That turns out to be the right incentive rather than a gap. A bare file with no explanation was never much use to the next reader anyway, agent or human. Making the write-up the index means the habit that makes research reusable is the same habit that makes it retrievable.

Which makes the workspace a research record

Do this for a few months and you have something a folder of downloads never becomes: an archive where every conclusion still points at the evidence that produced it. A new agent asked a question that's been looked at before finds the note, and one hop away, the workbook.

Keep the recurring work in a typed collection — one record per study, with the question, the date, the method and a link to the artifact — and the archive is queryable as well as browsable. Everything looking at pricing this year, newest first, with the models attached.