Other Blogs
Most of what I build is a variation of something I already built. A phone stand, then the same stand on a plate. A box, then a lid for that box. Until now the chat had no memory of your other models: you either described the first part all over again (and got a slightly different one back) or downloaded the file and uploaded it into the new chat.
Now you type @, pick the model, and the agent gets the real thing.

I had a phone stand from an earlier chat, 70 mm wide with a 12 mm lip. I wanted it on a square base with a cube next to it, mostly as a test of whether the agent would leave the stand alone.
In a new chat I typed "Put @", picked "Simple Phone Stand" from the list, and finished the sentence: "on a 100 mm square plate with a 20 mm tall cube beside it. Keep the stand exactly as it was."

The first thing the agent did was open the mentioned model. It got the stand's file, the OpenSCAD code that built it, and its seven parameters (width 70, lip 12, base depth 80 and so on). Then it wrote a Blender script that imports that file untouched and adds a 3 mm plate and the cube around it. Its reply: "the original stand is unchanged, placed on a 100 × 100 mm plate with a 20 mm cube beside it."
That's the point. It didn't build a new phone stand from the words "phone stand". It used mine.
Type @ anywhere in the message box. A list of your models opens, newest first, each with its thumbnail. Keep typing to filter by name ("@pho" is enough for "Simple Phone Stand"), move with the arrow keys, press Enter or Tab to insert. On a phone the list opens above the message box and you tap the model.
Two more ways in:

The list puts models you've mentioned before near the top, then models from the workspace you're in, then the rest by date. Delete the @Name text from your message and the mention goes with it, so the agent never gets a model you took out of the sentence.

A mention isn't a screenshot or a summary. When the agent opens a mentioned model it receives:
Opening a model costs nothing; it isn't a generation. The build that follows is a normal one.
What it does with all that depends on what you ask. "Keep it exactly as it was" means it imports the file and models around it. "Make the same thing 20% wider" means it adapts the code, because a mesh can't be stretched without ruining the details. "Match the hole pattern" means it reads the numbers out of the source. Among the steps in the chat you'll see "Opened @Name", and the chip links back to the original, so you can always check which one it used.
You can mention any model in a workspace you belong to, your teammates' models included, and any model that's been made public. Nothing else. A link to someone's private chat stays a plain link, and if the agent ever asks for a model you can't see, it gets a "not available" answer instead of the model.
Open a new chat, type @, and pick the last thing you printed. Ask for the lid, the wall mount, the two-up version, the one with your name on it.