Structured Memory Reduces LLM Cross-Domain Leakage and Sycophancy

Applying structured memory partitioning by domain during inference-time reduces cross-domain leakage and prevents model sycophancy. Testing across seven models on PersistBench shows that moving away from unstructured list-based memory injection improves response accuracy and topical boundaries.