Agent Memory Retrieval: Score-Space Landscape (2D)

This is the 2D counterpart to the 3D agent-memory simulator, and it uses a genuinely different geometry to show the same mechanism: instead of a 3D embedding sphere seen through an orbiting camera, memories live on a 1-parameter unit circle (an embedding angle) and are drawn twice — once on that raw circle, and once mapped into a similarity-versus-recency score-space where the retrieval formula itself is rendered as a shaded landscape. Reweighting semantic similarity, recency and importance visibly reshapes the landscape and reorders the ranked shortlist in real time, and as the bounded memory store fills up, the lowest-scoring memory against the current query is evicted — the same forgetting rule used by bounded, scored agent-memory systems like MemGPT-style tiered memory and time-weighted vector retrievers.