Agents talking to mcp-data-platform over MCP can capture a fact, a reviewer can promote it, and later sessions can search it. That knowledge system is the part of the platform I keep putting numbers on. The server is txn2/mcp-data-platform, Apache-2.0, also hosted as Plexara. The model stays fixed. One thing about the platform moves. Grading is deterministic. Every number recomputes from committed run data. The reports sit on Zenodo.

Five notes, in reading order: whether the knowledge layer makes an agent better on questions it cannot guess; when agents actually use a stored note, and on which model; the harness that runs the experiments; what a later agent does with a note that is wrong; and whether authored links between knowledge pages change what the agent can find, or only what it costs to find it.

A benchmark that can only agree with its author is marketing. This series is not a marketing page.