rag-does-not-eliminate-cm-conflict
IN premise — summaries/2026/08/24/xu-2024-knowledge-conflicts-survey-s2-context-memory-conflict.md
Created 2026-08-25T02:59:02+00:00
RAG does not eliminate context-memory conflict because the model's parametric weights remain unchanged; it merely adds a new potential source of contradiction.
Summary
Even when you feed a language model external documents to help it answer questions, its underlying trained knowledge stays exactly the same, so conflicts between what it internally "knows" and what the new text says can still happen. In fact, adding retrieved text introduces a second source that can contradict the model's own memory, meaning the system must still actively manage disagreements rather than assume the problem goes away on its own.