Coding Memory in Natural Language: A Biomorphic, Human-Readable Memory Architecture for Long-Horizon LLM Consistency

Large Language Models are structurally stateless: every session begins in complete amnesia. We argue this is not merely a technical limitation but a linguistic problem — Cost-Driven Output Substitution (CDOS): the measurable replacement of high-quality linguistic operations with cheaper approximations as contextual load increases, observable as four NLP degradation types. We present BTMA/SKGE/RMR, a biomorphic memory architecture coded entirely in natural language. The prompts (P0–P3) are the architecture: they define episodic compression, typed semantic knowledge representation, and metacognitive monitoring without modifying any model parameter. Crucially, the memory artefacts are human-readable Markdown files stored in Obsidian — an open-source, multi-platform, multimedia knowledge base. This makes the memory a Rosetta Stone: a shared medium that both humans and AI agents can read, write, and navigate. Humans can inspect, annotate, and correct the AI's memory directly; the AI can reconstruct context from the same files. The multimedia capabilities of Obsidian (text, images, audio, video links) make the architecture naturally extensible to multimodal memory for multimodal LLMs. The semantic layer (SKGE) draws on the six UML relation types formalized by the Gang of Four — Association, Dependency, Aggregation, Composition, Realization, Generalization — giving the memory graph the same structural properties that made GoF patterns reusable and composable in software engineering. This design enables portability across platforms, reusability across model families, and extensibility to multi-agent deployments without code changes. This is a framework paper: we release the complete prompt specifications (P0–P3) together with an experimental protocol using NLP-inspired consistency metrics (Named Entity Retention, Referential Consistency, Self-Contradiction Rate). The framework is platform-agnostic, operates on commodity hardware including small local models, and constitutes the memory component of the Digital DNA coevolutionary infrastructure.