Memory: See and Edit What Dagny Knows

There's a new Memory page, and it answers a question that used to be almost impossible to answer: what does my assistant actually know about us? Memory shows the full shelf of documents your assistant reads — the rules it follows, the memory docs for each area of your business, the knowledge and skills catalogs, and its activity log. Each document appears exactly as the assistant sees it, with no summarizing or reinterpreting in between. What you see is what it reads.

Reading is only half of it. Hit Edit on a document and it becomes plain markdown you can rewrite freely: fix a wrong detail, sharpen a line, move a fact into the section where it belongs. When you save, we reconcile your text back into the underlying records rather than replacing them wholesale — a line you didn't touch stays untouched, a line you move to another heading keeps its history and connections, and a line you reword updates in place. Deleting something is a deliberate act, so we ask you to confirm before anything is forgotten for good. If a save partially fails, the editor stays open with your work intact and retries only the part that didn't land.

Each memory document now follows a consistent scaffold — sections like Communication and Business basics, with gentle placeholders where the assistant hasn't learned anything yet. That makes the gaps visible, so you can fill them in yourself instead of waiting. Rules are folded into the same page, with full create, edit, and delete control returned to your team, and organized by section so a long list stays legible. Lessons your assistant has learned from your connected tools are editable too, right alongside the tools themselves — and they remain editable even after a tool is disconnected, so nothing gets stranded.

Getting there is simpler as well. "On Dagny's mind" and the separate Activity page have been retired in favor of clear heartbeat cards on Home, with old links redirecting automatically, so there's one obvious place for the live pulse and one for the long-term memory. We've also added guardrails throughout: clear limits with plain-language messages when a document or rule gets too long, a warning before you navigate away from unsaved edits, and a retry pane instead of a misleading empty page if something can't load.

The result is a fair bit of trust you didn't have before. When the assistant gets something wrong, you no longer have to coach it in circles — you can open the page, correct the sentence, and move on.