Competitive Docking Memory: Emergent Temporal Slot Specialization in Language Models

We introduce Competitive Docking Memory (CDM), a novel sequence modeling architecture that replaces attention with a bank of K learnable memory slots governed by independently trainable EMA decay rates. Tokens compete to write into slots via a load-balanced routing mechanism, and each slot's temporal horizon emerges from gradient descent. Without explicit supervision, CDM develops a multi-timescale temporal hierarchy: fast-decaying reactive slots specialize in volatile semantic content while persistent slots maintain structural context. At 85.7M parameters, CDM outperforms both a 72.9M standard transformer baseline and a parameter-matched 85.7M GQA+SwiGLU transformer (Δ−0.2688 nats, 15.5%). CDM V9, a dual-timescale HORN+Kuramoto variant, spontaneously self-organizes into functionally stratified routing circuits whose topology depends on training context length. A 15-measurement trajectory documents a six-phase coordinator lifecycle including Nash equilibrium convergence. CDM V10 extends findings to tri-timescale Basal Ganglia dynamics, confirming depth-dependent repeller dynamics across a four-way scale taxonomy. Code and model weights available on HuggingFace (DuoNeural).