From Retrieval to Faithful Memory Use: Context-Grounded Evidence Chains for Music Education Agents

Artificial intelligence systems are increasingly used for music education and music score understanding, yet most existing systems answer each task in isolation. This single-instance paradigm makes it difficult to evaluate whether an educational agent can accumulate experience, reuse prior solved cases, avoid repeated mistakes, and make its use of memory explicit. We propose a Memory-Tool Agent with Context-Grounded Evidence Chains for music education. The agent models memory as an agent-managed tool with explicit query, organization, and verification mechanisms, while supporting broader memory operations such as update and removal in future lifelong learning settings, and it is required to construct an explicit evidence chain over retrieved memory before producing an answer. We conduct experiments on MSU-Bench using a 100-example memory pool and a 1700-example test set with DeepSeek-v4-Flash. Compared with a BM25 retrieval-augmented generation baseline, the evidence-chain pipeline improves LLM-as-Judge accuracy from 56.6% to 57.3%, raises explicit memory use to 99.6%, and increases average retrieved-memory usage from 1.5% to 75.4%. Furthermore, the non-dependence rate (answers indistinguishable from no-memory generation) is reduced from 54.7% to 8.1%, and our method demonstrates strong robustness against counterfactual memory poisoning, with only a 1.0% accuracy drop under a single poisoned memory. These results underscore the importance of explicit, verifiable memory-grounding mechanisms for building faithful AI music education agents. In the current benchmark setting, memory is initialized from previously solved examples, providing a controlled evaluation of memory-grounded reasoning rather than a full lifelong learning scenario.