factwashing

This paper investigates how epistemic stance (qualifiers, attributions) survives memory compression in AI agent memory systems. It finds that making the stance explicit as a labelled field improves retention significantly, while merely lengthening the text does not.

Introduces factwash, an open-source write-time gate that deterministically catches AI rewrites that turn hearsay into fact by stripping attribution, hedging, or temporal context. The paper analyzes when simple checks suffice versus when an LLM witness is needed, and evaluates the approach on large annotated corpora.