Systems that retrieve from records which revise themselves, issue threads, policy logs, encyclopedic histories, long conversations, face a harder question than finding relevant text. They have to judge which claims still hold, which were superseded, and when the honest answer is to decline.

Structured memory is the usual proposal: typed edges, temporal updates, explicit conflict status. The problem this paper identifies is in how those proposals get evaluated. Studies tend to change the underlying mechanism and the way information is presented in the prompt at the same time, so a measured improvement cannot be attributed to either.

They call it a render confound, and it is the kind of methodological finding that quietly invalidates a body of results. If reformatting the prompt would have produced the same gain, then the elaborate memory structure was never the thing doing the work, and nobody could tell because the two moved together.