Abstract Large language models are commonly described as operating “in context,” yet what they actually consume is a finite, sliding window of tokens. This string-based mechanism bears some semblance to how humans maintain and use context-through concurrent, object-based, and selectively governed thought - but the similarity is superficial. As a result, human-AI collaboration is plagued not primarily by forgetting, but by misalignment of attention: some important ideas fall out of the token window while irrelevant ones persist, forcing the model to infer what the human still cares about from a decaying transcript. By reframing context as a curated collection of identity-stable, time-anchored objects rather than a linear string, Cognitive Memoisation enables reliable continuity, prevents semantic drift, and makes distributed cognition computationally tractable.-- This is an anchored non-peer reviewed paper