Distil the Competence, Externalise the Memory: Where Should an Embodied Agent's Experience Live?

Where should an embodied agent's accumulated experience live --- folded into its policy weights, or kept external and retrieved at inference? We study this under pre-registration on a vision-conditioned reach task. Distilling a memory-augmented teacher into student weights did not carry the memory benefit ($+1$ pp over an action-only control, n.s.); attaching the \emph{same} memory as an inference time retrieval context to a \emph{fixed} set of student weights recovered it ($+34$ pp on identical weights, $z{=}5.15$, exploratory; $+23$ pp on the pre-registered protocol contrast, $z{=}3.36$, confirmatory), matching the teacher's 49\% success at substantially lower inference cost. A pre-registered probe localises the mechanism: distillation transfers the \emph{marginal} correction (a static prior) but not the \emph{conditional}, situation-specific one that retrieval supplies. On a second, harder task the effect points the same direction but is not significant ($+6$ pp, sub-MDE). The recipe: distil the competence, externalise the memory.