SAVANT: Sparse Adaptive Vector Alignment for Continual Learning and Persistent Agent Safety

Official Preprint: CALMS-SAVANT Architecture Persistent LLM agents suffer from a critical vulnerability during continual learning: catastrophic consolidation. Under longitudinal adversarial pressure, standard replay mechanisms passively assimilate stealth poisoning, leading to permanent alignment collapse. This paper introduces CALMS-SAVANT, an asynchronously regulated plasticity architecture that solves this by replacing unrestricted parameter updates with rigorous geometric validation and Byzantine-robust initialization. Key Highlights: Multi-Seed Consensus Prior: Eliminates epoch-0 vulnerabilities by constructing the initial Identity Graph via Byzantine-fault-tolerant geometric arbitration. Geometric Stability Filter: Quantifies global manifold drift using Sinkhorn-regularized Wasserstein-1 distance. The Cognitive Diode: Executes adaptive gradient suppression to block adversarial parametric integration. PPP-Bench: Introduces the Persistent Prompt Poisoning Benchmark for longitudinal evaluation over 100 epochs. Results: Evaluated on LLaMA-3-8B, CALMS-SAVANT reduced the terminal Poison Success Rate (PSR) to 4.8% (down from 89.7% in standard replay), while safely preserving a Fact Retention Score (FRS) of 92.3%. Code and benchmark datasets are being prepared for release on GitHub.