We present Codette, a modular cognitive architecture that models multi-perspective reasoning as a constrained dynamical system converging toward stable cognitive attractors under explicit modeling assumptions. The system integrates six heterogeneous reasoning agents (analytical, creative, ethical, philosophical, probabilistic reasoning, and empathic), a persistent memory substrate (cocoons), and a meta-cognitive engine that discovers cross-domain reasoning patterns and generates novel reasoning strategies from its own history. The RC+\(\xi\) (Recursive Convergence + Epistemic Tension) formalism provides a dynamical-systems-inspired lens for describing cognitive state evolution; convergence is conditional on modeling assumptions detailed in Sect. 3 and is not claimed as a general guarantee. We evaluate Codette through an ablation-style benchmark suite of 17 problems across six categories under four conditions: single-agent baseline, multi-perspective synthesis, memory-augmented reasoning, and full Codette with strategy evolution. On the May 2026 benchmark run (951 stored cocoons), the full system achieves a composite quality score of 0.744 vs. the single-agent baseline of 0.357 (+108.8%, Cohen’s \(d=8.31\)). Gains on length-sensitive dimensions (perspective diversity, reasoning depth) are by architectural design; length-independent dimensions (coherence, factual grounding, Turing naturalness) show smaller, more modest improvements discussed in “Response length analysis”. Memory augmentation reaches statistical significance at this scale (\(p=0.020\), \(d=0.80\)), resolving a prior null result at smaller scale (217 cocoons). An independent external evaluation on GPQA Diamond (\(N=198\) graduate-level questions) provides out-of-distribution corroboration. The architecture runs on consumer hardware (Llama 3.1 8B with ten LoRA adapters) and is open-source (Zenodo DOI: https://doi.org/10.5281/zenodo.19359663).
