📝 SummarXiv

Abstract

How much energy must an embodied agent spend to remember its past actions? We present Tangential Action Spaces (TAS), a differential-geometric framework revealing a fundamental trade-off between memory and energy in embodied agents. By modeling agents as hierarchical manifolds with projections Phi: P -> C and Psi: C -> I connecting physical (P), cognitive (C), and intentional (I) spaces, we show that the geometry of Phi dictates both memory mechanisms and their energetic costs. Our main contributions are: (1) a rigorous classification proving that one-to-one projections (diffeomorphisms) require engineered dynamics for memory while many-to-one projections (fibrations) enable intrinsic geometric memory through connection curvature; (2) a proof that any deviation from the energy-minimal lift incurs a quantifiable penalty, establishing that path-dependent behavior necessarily costs energy; and (3) a universal principle that excess cost Delta E scales with the square of accumulated holonomy (geometric memory). We validate this cost-memory duality through five systems: the strip-sine system (engineered memory, Delta E proportional to (Delta h)^2), helical and twisted fibrations (intrinsic geometric memory), and flat/cylindrical fibrations (proving curvature, not topology, creates memory). This framework bridges geometric mechanics and embodied cognition, explaining biological motor diversity and providing design principles for efficient robotic control.