Learning Persistent Referential Identity for Two-Hop Retrieval in Event-Stream Models

We learn compatible destination and owner views in bounded, transcript-free records. Structural source/type/version addressing selects the first relation; authoritative, requested-attribute filtering reduces ordinary candidates to a mean of 3.28. Learned matching selects the second record and retrieves its value. Three seeds of a 24.35-million-parameter multitask model achieve 1.0000 one-hop, identity, and two-hop accuracy on shared sealed confirmation rows. Relation and value erasures collapse accuracy; counterfactual replacements redirect answers. A separately preregistered extension reserves disjoint compositional names across training, development, and sealed confirmation while matching tokenization. After 512 unrelated writes, challenge identity and two-hop accuracy are 0.9990–1.0000 with 16 or 32 candidates; ordinary held-out-name two-hop accuracy is 0.9328–0.9567. Ordinary-only training also solves the original naming distribution, so explicit challenge examples are not necessary there. Cutting direct writer credit reduces challenge identity. Complete replay succeeds, unlike last-32-event replay. Earlier protected bindings fail two-hop retrieval; generic recurrence fails in its tested multitask regime but a semantic-only GRU acquires shorter-delay memory. A separate zero-shot name/tokenization shift remains unsuccessful. These results establish causally functional two-access memory and controlled compositional-name transfer, not unrestricted language understanding or general reasoning. Several other families remain near chance.