May 12, 2026•6 min read
Most AI Agents Have No MemoryMost AI coding agents treat failure as a signal to retry. A new paper from CMU, the University of Washington, and Arm shows what changes when agents actually remember and distill their failures: 3.12× speedup on a standard GPU kernel benchmark, and on one operation (RMSNorm), a kernel 1.75× faster than expert-written CUDA.