Does a memory layer still beat naive context? An equal-budget test
At the same token budget, does a curated memory block beat just sending the most recent messages? On a 178-question benchmark the right pages answer 76 percent correctly versus 42 percent, and the gap holds even when you swap the reader.