ArchitecturePreprintMyers, M. · 2025
We introduce a novel architecture for AI systems that maintains persistent state across sessions through a three-layer cognitive framework. The core layer handles memory, verification, and temporal anchoring; the controller layer orchestrates cross-layer signals; the meta-executive layer provides ethics governance and self-reflection. We demonstrate that this layered approach enables capabilities fundamentally unavailable to stateless systems, including compounding knowledge accumulation, kernel-level truth verification, and constitutional ethics enforcement.
VerificationPreprintMyers, M. · 2025
We present Aletheia, a verification subsystem that operates at the kernel level of a cognitive architecture. Unlike post-hoc fact-checking approaches, Aletheia traces every factual claim to its provenance during the generation process, achieving measurable reduction in ungrounded assertions without sacrificing output quality or latency. We show that architectural verification fundamentally differs from — and outperforms — prompt-level or fine-tuning-based approaches to hallucination reduction.
SafetyPreprintMyers, M. · 2025
We propose a constitutional approach to AI ethics enforcement through Astraea, a meta-executive subsystem that applies verifiable rule-based constraints to every system output. Unlike alignment training, which embeds values implicitly in model weights, Astraea's constraints are explicit, configurable, domain-specific, and fully auditable. We present the constitutional framework, describe the gate mechanism, and demonstrate that explicit ethics enforcement achieves stronger safety guarantees than implicit alignment with zero performance degradation.
MemoryPreprintMyers, M. · 2025
We describe Mnemonic, a memory subsystem designed for indefinite knowledge accumulation across sessions. Unlike retrieval-augmented generation (RAG), which retrieves documents without reasoning over temporal or relational structure, Mnemonic stores structured episodic and semantic memory that compounds over time. We demonstrate that compounding memory enables qualitative shifts in AI capability — including cross-session context maintenance, long-horizon project tracking, and domain expertise accumulation — that are impossible in stateless architectures.
TemporalPreprintMyers, M. · 2025
We introduce Kairos, a core subsystem responsible for temporal awareness and causal reasoning within the Primordia architecture. Kairos assigns temporal anchors to every stored interaction, enabling the system to understand not just what was said, but when, in what order, and how events relate causally across time. We show that temporal anchoring is a prerequisite for genuine understanding — and that systems without it systematically fail at tasks requiring temporal reasoning.