It's the persistent structure that retains context between interactions, outside the model. The MIT study explicitly recommends funding it: if it doesn't retain context, it doesn't scale.
Why outside the model, not inside it?
Because knowledge embedded in the model disappears when the model is swapped, can't be audited and can't be removed. A separate structure is cheaper, portable and verifiable.
Is this the same as a document database?
No. A document holds what was written; the layer holds the relationship between facts — who depends on whom, where practice diverges from the record — with the source tagged on every end.
Who queries this layer afterwards?
Whoever decides, through the dashboard, and any of the company's agents through an open protocol, with the same access control. The layer isn't a separate product: it's the context everything else now has.
Fine-tuning embeds knowledge in the model and loses it at the next swap. The layer does the opposite: it outlives the model that queries it.