The paper studies how to prevent interference when LLM agents store persistent skill edits across recurring task families. It introduces family-scoped retrieval, which limits retrieval of accepted skills to the family that produced them.
In experiments on a 12-round code-repair stream (ProcStream-RSI), this scoping increased mean hidden trajectory utility from 0.713 under global memory to 0.816, and reduced harmful deployments from six of eight to none. In 27 randomized-order streams, scoped retrieval improved utility by 0.063 over global retrieval and accepted 63 updates (vs 12) with zero harmful.
These results show that persistent improvement requires not just certifying an edit as valid, but also matching its deployment scope to its evidence scope.