Metamemory-guided retrieval and memory revision for long-term conversational memory

Relevant information in long-term conversations is often scattered across sessions. Even when a system finds evidence that supports its answer, it may omit other information needed to address the question. To address this problem, we propose a framework for metamemory-guided active retrieval and memory revision (MARR). The framework organizes historical conversations into a heterogeneous memory graph, connecting records across sessions through entities and topics while retaining original messages and temporal information for later inspection. When answering a question, the system checks whether the evidence is sufficient in light of what it has already answered and uses the actual retrieval trajectory to determine where to search next. Information obtained through supplementary retrieval informs revision proposals, which the Writer checks against the original text before deciding whether to accept them, reducing unsupported changes. Evaluations on LoCoMo and LongMemEval-S show that the framework recovers some omitted evidence and improves question-answering performance. Further analysis finds that some answer errors can be corrected by reexamining material already available, while continued retrieval yields more pronounced benefits for multi-hop questions with incomplete evidence. These findings suggest that the effectiveness of long-term conversational memory depends both on finding relevant information and on the system's ability to understand and use it adequately.