Abstract Reasoning in complex, poorly understood domains requires methods that tolerate incompleteness while still preserving progress. Large Language Models (LLMs) are effective exploratory tools, but their stateless and non-authoritative nature leads to repetitive rediscovery and loss of hard-won understanding. This paper introduces Cognitive Memoisation as a lightweight knowledge-engineering pattern that externalises facts, constraints, and invariants into human-governed artefacts. When paired with LLMs, Cognitive Memoisation enables sustained exploratory modelling, prevents Groundhog Day–style rediscovery, and provides a disciplined bridge from early abstraction to later formalisation.