Marginal Advantage Accumulation for Memory-Driven Agent Self-Evolution

arXiv·medium signal

Proposes MAA, which fixes contradictory cross-batch feedback in trace distillation by building differential signals, accumulating per-operation evidence via EMA, and merging semantic identities for traceability. Reports superior performance on 14 of 16 test settings while cutting token consumption ~75% vs existing methods. Directly relevant to anyone building self-improving agent memory.