Verification-Gated Persona State Transitions for Memory-Augmented Language Agents

Memory-augmented language agents are increasingly used for long-horizon personalized interaction, but they still struggle to adapt to user-specific evidence without gradually drifting away from a stable and reliable persona. In this paper, we study how to maintain long-term persona invariance while preserving the utility of memory-based personalization. We propose a verification-aware memory framework that separates fast-changing episodic memory from a slow persona state and treats persona evolution as a controlled state transition rather than an unconstrained memory write. At a high level, retrieved evidence can only induce a bounded tentative persona update, and the update is committed only if an external verifier confirms that the resulting symbolic trace satisfies explicit persona specifications. Across long-term memory, personalized dialogue, persona consistency, and over-personalization benchmarks, our method achieves the best overall performance, improving utility metrics on LongMemEval, LoCoMo, and PersonaMem-v2 while reducing hard violation rate from 6.9 to 3.2 and increasing repair success rate from 67.5 to 74.9 on reliability benchmarks. These results show that explicit verification can make memory-augmented agents both more adaptive and more dependable, offering a practical path toward safer long-term personalization.