MaSRead: Content-Addressed Reading for Replicated Latent Stores

A recent preprint on arXiv (2608.11218) presents MaSRead, a technique designed for content-addressed reading of replicated latent stores. This system allows independent agents operating in latent space to exchange computed states as key-value cache fragments instead of plain text. These fragments are combined using a conflict-free replicated data type, which guarantees convergence regardless of the order or duplication of delivery. Nonetheless, a subsequent query that is not known at the time of encoding cannot accurately access the merged cache due to interference from colocated fragments, rendering colocation inadequate for addressability. MaSRead resolves this issue by routing through opaque keyed tag sets based on fragment words and decoding each chosen fragment with a hard attention mask that conceals the remainder. Utilizing lexical connectivity, a graph traversal accesses the necessary fragments for multi-hop queries. The method was evaluated on various types of stores—including chain, pipeline, symmetric, hub, and natural-language stores—successfully retrieving isolated visited fragments, maintaining effectiveness as unrelated fragments accumulate, and adapting to a different model family. The paper is available on arXiv with the identifier 2608.11218.