This work addresses adaptive recommendation under intent drift. Feedback from each recommendation reveals whether the relational evidence used for ranking is useful, missing, or misleading. Knowledge Graphs (KG) supply essential semantic structure, yet conventional KG‑enhanced systems treat the graph as a static retrieval substrate, making them brittle to evolving intents, noisy metadata, and recurring failure patterns. CGM‑Rec is introduced as a continual graph memory framework. It treats the graph state as writable memory and maintains two complementary components.
The Semantic Graph Memory is updated conservatively via quality‑gated typed operations to store stable, high‑confidence relational knowledge. The Episodic Lesson Memory serves as a fast reactive memory that learns recent outcomes, failure cases, and corrective hints. During testing, model parameters stay frozen; adaptation occurs solely through memory writes.
Evaluation follows a frozen‑parameter, one‑pass reranking protocol where encoders and prompts remain fixed and all adaptation is performed by writing to memory. Across multiple recommendation settings, CGM‑Rec outperforms neural and LLM baselines on most metrics. Notably, under sampled‑candidate reranking, CGM‑Rec improves HR@1 by up to 29.58% over the strongest LLM baseline on the Bundle dataset, and on metadata‑rich ML‑100K it achieves HR@5 of 0.5941 versus 0.4746 for K‑RagRec.
Review: CGM‑Rec converts a knowledge graph into a writable memory, enabling rapid adaptation to intent drift without altering model parameters, and highlights the promise of memory‑augmented recommendation systems.