I’m the author of this paper. I've been researching cognitive architectures and found that while LLMs are scaling impressively, they are hitting architectural ceilings regarding persistent memory, reasoning depth ("System 2"), and energy efficiency.

I wrote this paper to propose MQ-AGI, a theoretical framework that attempts to solve these bottlenecks through Orchestrated Modularity rather than just parameter scaling.

The core concepts:

Decomposition: Instead of a monolith, the system uses specialized Domain Expert Networks (DENs) coordinated by a Global Integrator Network (GIN) — heavily inspired by Global Workspace Theory.

Quantum-Inspired Routing: The "secret sauce" is treating the selection of experts not as a statistical gating problem (like in MoEs), but as a Combinatorial Optimization problem. I model this using Hamiltonian energy minimization to find the optimal coalition of experts for complex tasks.

Note: I explicitly address the QRAM bottleneck and latency issues in the paper. The proposal focuses on optimizing low-dimensional metadata (routing signals), making it feasible for NISQ devices or classical simulation via Tensor Networks.

DREAM Memory: It integrates a hierarchical memory protocol (Episodic/Semantic) with adaptive retention (TTL) based on user engagement, rather than raw context window stuffing.

The paper includes the full mathematical formalization (Hamiltonians, Free Energy Principle equations) and a critical feasibility analysis.

It’s a preprint and a conceptual blueprint. I’d love to hear your feedback on the routing topology and the feasibility of the hybrid core. https://zenodo.org/records/17654543