Optimising Memory Storage and Retrieval in Long-Term Counselling Dialogue

​

Large language model (LLM)-based systems are increasingly used for mental-health counselling, where continuity of care depends on remembering a client's past disclosures. As dialogues grow, however, conversation logs become massive and counselling-critical information—such as suicidal ideation or life turning points—gets buried, reducing retrieval accuracy. Existing long-term memory methods (SeCom, LUFY) are not tailored to the counselling domain: they treat casual chat and clinically important utterances equally, or judge importance by general emotional arousal, discarding low-arousal "quiet SOS" signals. This study proposes Adaptive Indexing with Selective Raw Response. Dialogue is managed in topic-based segments, each holding the raw log, a summary, and structured KeyPoints across five counselling categories (Risks, Turning Points, Stressors, Actions, Supports). A Raw Preservation Score (RPS) quantifies clinical importance, and the search index is built adaptively—retaining raw text only for high-importance segments to reduce noise. At generation time, raw utterances are presented only for risky or high-RPS segments. On a 10-session, roughly 300-utterance counselling scenario, the method reached 0.767 on Complex QA, exceeding SeCom (0.733), and raised Recall@5 from 44.1% to 62.0% while cutting raw-text volume by about 55%. The results show that importance-aware adaptive memory management improves retrieval of counselling-critical memories.