Sinapsi: Separating Fluid and Crystallized Intelligence: a Five-Lobe Topology for Continuous Latent Reasoning

Abstract

In current language models the ability to reason (fluid intelligence) and knowledge about the world (crystallized intelligence) occupy the same parameters, and explicit reasoning must pass one step at a time through a single token. Sinapsi is a modular five-lobe architecture that separates these two components by construction. A recurrent inference engine, holding no facts, reasons in a continuous latent space; knowledge lives in an external vector memory (the Hippocampus) compiled by the model’s own encoder, both offline and at run time, so that facts reach the engine as native vectors rather than as text; language appears only in the peripheral lobes, and the decoder may copy from the facts only the fragments selected upstream. A passive observer, the Watchdog, writes the path of the reasoning into a read-only register that the engine consults to keep its bearings, together with an access log that makes what the engine consulted verifiable. The underlying thesis is a reallocation of the parameter budget: what is not spent on remembering can be saved, yielding smaller models, or invested in reasoning capacity, either by enlarging the engine or by composing many specialized Sinapsi into a network that shares latent space and memory. The work states its claims as falsifiable hypotheses and proposes a minimal experiment, the Knowledge Swap Test, that runs on a single consumer GPU.