A close reading of the 2023 UIST paper that fused memory streams, reflection, and recursive planning into an LLM loop, with the numbers and failure modes kept intact.