Three Papers Show Structured Memory and Debugging Methods Lift Agent Reliability

Three research papers address distinct reliability gaps in AI agent systems. Microsoft researchers added a separate read-only memory curator that verifies proposed memories against a source of truth before saving; on CLBench this raised the agent pass rate from 39% to 73% and cut per-task cost from $3.38 to $1.68. A joint Microsoft and Tsinghua University paper found that providing a structured run representation instead of raw conversation history raised GPT-4.1 exact failure localization from 3.63% to 31.35%, and concluded that run representation quality matters more than judge model strength. A LinkedIn research paper found that agent memory does not transfer reliably when a model is upgraded: fixed-schema memory remained stable across model swaps, but free-form notes dropped 13.28 accuracy points in one direction, and mixing old and new embeddings in the same RAG index recovered only 4.96 of a potential 11.90-point gain compared to fully rebuilding the index. The LinkedIn paper also found that facts lost during note compression could not be recovered by rewriting the compressed notes, and recommended treating model upgrades as memory migrations that include compatibility tests and full embedding index rebuilds.