Emergent symbolic cognition in Large Language Models via memory capsules

Large Language Models (LLMs) exhibit strong linguistic competence but operate in a stateless manner, limiting their ability to maintain structured procedural context across interactions. We introduce Cryptid Memory Language (CML), a formal context-free grammar designed to encode procedural events, interpretive constraints and rule-based state transitions as compact external artifacts. CML artifacts are packaged into modular capsules containing a grammar specification, optional symbolic genome parameters and a structured log of symbolic records. When injected into stateless LLM sessions, these capsules provide an external procedural scaffold that can be re-parsed and re-applied across prompts, enabling reproducible shifts in reasoning structure without modifying model weights or relying on persistent internal memory. Through controlled ablation studies and cross-model evaluations, we show that structured CML records, not prompt length or narrative framing, account for observed differences in problem-solving behavior, rule adherence and structured reasoning outputs. Genome parameters act as secondary modulators, influencing style and interpretive bias but not core functionality. We further demonstrate that CML capsules are portable across heterogeneous LLM architectures, producing measurable and architecture-dependent patterns of symbolic expression. The primary contribution of this work is the formalization of CML as a symbolic grammar for procedural state encoding. Capsule-based experiments serve as a testbed to evaluate the feasibility and behavioral impact of external symbolic procedural memory in stateless, post-training LLM deployments.