Agents stop writing every lesson to memory: FGSE tests each update for transfer, falsification and regression first

Agent self-evolution moves from writing experience directly into memory to treating each experience as a hypothesis under test: a behavioral update distilled from an episode enters persistent memory only after passing transfer, falsification and regression checks. Earlier self-feedback, reflection, skill-memory and context-evolution methods typically write episode lessons into memory or rewrite context directly, without a criterion for an update's transferability, retention value or commitment risk. Across verifier-rich benchmarks including WebArena, AppWorld, τ-bench, AgentDojo, LongMemEval and LoCoMo, FGSE is compared with reflection-, memory- and context-evolution baselines and improves task performance while reducing falsification failures, regression damage and harmful committed updates.

Each candidate update is written as a structured hypothesis with four parts, a precondition, a behavior transformation, an expected effect and a verifier, and it acts only in a temporary state without immediately changing persistent memory. Free-form lessons tend to be too broad; the diagnostic study shows that broad, underspecified lessons cause larger boundary and archive harm than specific ones, and sequential writing accumulates errors. The commitment gate judges on three probe statistics, target-transfer gain, falsification failure rate and regression damage, which respectively cover cases the hypothesis should help, cases designed to expose overgeneralization, and cases the agent already solves and must not degrade; probes are admitted only when an executable or externally checkable signal can score them.

A hypothesis that fails the gate but whose failures fall in a separable sub-scope gets one precondition-narrowing pass and is re-evaluated; one that still fails is rejected, and committed hypotheses are archived with their probes, verifier outcomes and post-commit violation record. Ablation shows the largest drop comes from removing structured hypotheses, with WebArena success rate down 4.2, AppWorld average completion down 5.9 and LoCoMo average R-L down 4.78; removing falsification probes raises falsification failure from 5.0% to 18.6% and harmful updates from 3.6% to 10.6%. At runtime only committed hypotheses whose preconditions match and whose violation record stays below threshold are retrieved; multiple active hypotheses are combined by normalized precondition scores while preserving a nonzero base-policy contribution, and hypotheses exceeding the violation threshold are removed from retrieval.

The demonstrated scope is verifier-rich settings: web and browser control, app and API interaction, tool use, adversarial tool outputs and long-term memory benchmarks, plus coding and mathematical reasoning diagnostics; all experiments use frozen language models with no training or fine-tuning, changing only persistent agent state and the commitment rule. The authors propose extending falsification-gated control to multimodal agents, collaborative tool-use systems, and settings where verifiers must be learned or approximated.

The gate's theoretical guarantee is relative to the probe-generation distributions, and a deployment distribution that differs from them requires an additional distribution-shift term; the quality of the probe generator determines how well those distributions cover deployment-time boundary cases. The ablation also shows that repository-level changes such as SWE-bench remain regression-prone because they are highly contextual.