MaSRead lets AI agents read each other's latent caches without interference

Independent AI agents that reason in latent space often compute shared state—key-value cache fragments—that other agents could reuse. But merging these fragments into a replicated store using conflict-free replicated data types (CRDTs) creates a problem: when a later query arrives, colocated fragments interfere with each other, so colocation doesn't guarantee addressability. MaSRead, a new system from Carlos Baquero, Luís Brito, and João Resende, solves this by routing reads based on content, rather than location.

MaSRead derives opaque keyed tag sets from fragment words and decodes each selected fragment under a hard attention mask that hides the rest of the store. A graph walk under lexical connectivity reaches fragments required for multi-hop queries. Experiments across chain, pipeline, symmetric, hub, and natural-language stores show MaSRead recovers visited fragments in isolation, stays effective as unrelated fragments accumulate, and transfers to another model family. Materialized decoding depends on fragment length rather than total store size, though end-to-end work still includes store-dependent routing and one read per visited fragment. The approach has known limits: lexical routing can miss disconnected evidence, and answer composition is bounded by the frozen reader. Still, replicated latent stores become selectively readable when fragments connect to the query through content.