Retrieval, not recall
An assistant that looks things up
beats one that remembers.
Every language model carries a fixed snapshot of the world, frozen on the day
its training stopped. It will answer confidently from that snapshot long after
reality has moved on — and it has no way to notice that it has, because stale
knowledge and current knowledge feel identical from the inside.
So our agents are built to retrieve before they answer. Ask
one about your project and it reads your project, not its memory of projects.
Ask it about a tool and it consults that tool's current documentation. What it
recalls is treated as a hint; what it retrieves is treated as evidence.
This is not a performance optimisation. It is what makes the rest of our
governance possible: an agent that retrieves can cite, and an agent
that cites can be checked. One that merely recalls can only assert,
and asking a person to approve an assertion they cannot verify is not
oversight — it is a signature on a blank page.
The same discipline is why this company's AI-assisted work is written to
verify against live sources rather than trust a summary of them. We have
watched it catch real errors — including our own, including on this site.