A Past Crisis, a Present Reply: Learning to Regulate Memory Influence

Conversational memory helps language models maintain continuity in supportive dialogue, but salient past experiences can also draw responses away from a user's current needs. Historical information may remain valuable for factual access even when it is no longer relevant to the present request, creating a need to regulate memory influence without removing the memory itself. We introduce Memory Influence Budget (MIB), a fine-tuning framework that separates memory access from memory influence by placing a calibrated budget on changes in response distributions under matched memory interventions. Experiments demonstrate that MIB substantially reduces historical influence without a corresponding loss of memory dependent capability. Across model backbones and public memory QA tasks, MIB preserves factual access while maintaining current evidence recognition. Comparisons across control strengths further demonstrate stronger factual access than reverse context-aware decoding at comparable influence levels. These results demonstrate separate control of memory access and influence, enabling reduced historical influence without sacrificing access to retained information.