We present Noemica, a cognitive architecture where thought representations are language-independent latent states, stored in compressed form and expanded only on task demand. Through cross-lingual hidden state analysis of Gemma 4 12B and Qwen 3.5 9B, we identify a Noema Point — a layer range where representations converge across languages regardless of language family. In Gemma 4, cross-lingual cosine similarity reaches 0.98-0.99 at layers 14-17 with a sharp phase transition (ignition) at layer 13. Convergence holds across six languages from four families (Indo-European, Sino-Tibetan, Semitic, Dravidian) with IE vs non-IE difference of only 0.004. Direct thought transfer between agents achieves 22,786x lower semantic loss than language-mediated communication. The Noemica architecture comprises three cognitive pillars (perception, reasoning, dream), an associative thought graph with task-driven lazy expansion, and a Mixture-of-Experts verbalization layer. This provides a path beyond token-prediction toward language-independent machine cognition.