Reproduce “Interactive Memory Learning for Long-Term Conversations” on OpenResearch

About this paper

Cai Ke, Jiangyue Yan, Han Zhang, Xin Liu, Zike Yuan, Yue Yu, Hui Wang, Ruifeng Xu

Interactive Memory Learning for Long-Term Conversations models memory management as an adaptive sequential policy rather than a static archive. A Planner learns what to store and a Trigger learns when to retrieve it, with delayed cross-session rewards linking earlier storage decisions to later response quality; across three multi-session benchmarks, the system improves automatic and human-rated dialogue quality while reducing memory-selection errors over time.

Key insights

• Memory value is best evaluated by its later contribution to a conversation, not only by its intrinsic content or semantic similarity to a query.
• Storage and retrieval are coupled decisions: a memory is useful only if it is retained appropriately and retrieved at the right time.
• Online adaptation steadily improves policy behavior; the reported high-value-memory miss rate falls from 48.2% initially to 3.2% by the fifth session, while false positives fall from 15.6% to 6.1%.

Summary by alphaXiv.