LLM agents are increasingly deployed in long‑horizon, verifier‑rich settings. Existing self‑improvement loops often write episode‑derived lessons straight into memory, so local success is not a reliable indicator of reusable and safe knowledge. This work tackles the fundamental question: when should an experience‑derived behavioral update be admitted to persistent state?
We introduce Falsification‑Gated Self‑Evolution (FGSE), a verifier‑grounded framework that encodes each candidate lesson as a structured hypothesis comprising a precondition, behavior change, expected effect, and verifier. FGSE first applies the hypothesis in a temporary state, then probes it with target‑transfer, falsification, and archived‑regression tests. Based on measured gain and risk, the hypothesis is committed, refined, or rejected.
Across WebArena, AppWorld, tool‑use, and long‑term memory benchmarks, FGSE achieves strong performance: 39.8% success on WebArena, 78.8% average completion on AppWorld, and Pass$^1$ scores of 0.724 / 0.487 on (\tau)-Retail / (\tau)-Airline. At the same time, falsification failures drop to 5.0%, regression damage to 1.3%, and harmful committed updates to 3.6%. These results suggest that testing memory updates as hypotheses supports more reliable long‑term adaptation.
Review: FGSE offers a systematic, safety‑oriented approach to memory management for LLM agents, balancing learning efficiency with risk mitigation.