The Memory Tax: Smart Models Don't Need the Scaffolding
A new study tested seven memory approaches on 152 long-conversation questions. Weaker models gain from fact-extraction memory, while strong models do as well or better with raw history. The flagship model scored about 95% either way.
The most striking finding is that whoever writes the memory sets the ceiling. A knowledge graph built by a small model capped a frontier reader at 45%. When the frontier model built the graph itself, it scored 87%. A lightweight harness reached 99% for $6.95, against $17.84 for stuffing the full context. A popular memory layer breaks even only after 372 to 596 questions.