The article frames retrieval-augmented generation as a high-overhead translation layer rather than a true memory system. Current pipelines compress a model's hidden states into text, re-encode them via a separate embedding model into vectors, store them in a vector database, retrieve matches, then reconstruct a prompt for a third model to rebuild hidden states from scratch — a seven-stage chain where only two stages are neural-native. The other five exist solely because models cannot yet persist their own state directly.

Larger context windows do not solve the core problem. A two-million-token window still fails at portability and persistence when autonomous agents hand off tasks, edge devices move between compute clusters, or multi-agent pipelines span different machines. Transferring a full prompt incurs punishing bandwidth costs and forces the receiver to re-read every token — a full prefill pass — to reconstruct reasoning state.

A back-of-the-envelope latency budget shows a single RAG call consuming roughly 135 ms across sequential steps: token generation, embedding, network I/O, vector search, reranking, prompt reconstruction, and decoding. In chat this is invisible; in robotics control loops, haptic feedback, or self-driving stacks, that entire budget is spent on plumbing before the model produces useful output. Direct GPU-to-GPU latent state transfer eliminates the embedding, network hop, retrieval, reranking, and reconstruction steps entirely.

The field has hollowed out translation layers before: raw files, relational databases, search indices, text embeddings, vector search — each became a substrate, not the primary interface. Latent persistence is next, but interoperability is spectactularly hard: architectural compatibility, precision matching, layer normalization scales, and rotary embedding alignment must all align, or transferred state decodes as confident nonsense.