Triadic Memory Governors for Deterministic LLM

Large Language Models (LLMs) exhibit impressive reasoning capabilities but suffer from long-context instability, memory drift, and run-to-run variance in extended interactions. Common mitigations (retrieval-augmented generation, heuristic memory buffers) remain partly stochastic and weakly interpretable at the level required for regulated or safety-critical deployments. We propose a deterministic external control architecture, termed a Triadic Memory Governor, that acts as a non-Markovian state observer and policy layer around an otherwise unchanged LLM. The governor embeds interaction history into a triadic state ψ(t) = (t, ϕ(t), χ(t)), where ϕ(t) operationalizes contextual alignment and χ(t) := ∂tϕ(t) captures memory torsion and regime transitions. A preregistered instability functional aggregates retrieval deviation, information inconsistency, cross-context coupling drift, and torsion into a transparent control signal that deterministically gates retrieval, verification, and memory writes. The framework does not modify model weights, training procedures, or proprietary internals. It provides a reproducible, auditable memory-governance layer suitable for agent systems, decision support, and regulated AI applications.