ConvMemory v2: A Recall-Preserving Top-10 Evidence Reranker for Conversational Memory Retrieval

A new conversational memory retrieval model, ConvMemory v2, improves ranking precision over its predecessor while preserving recall, according to a technical report posted to arXiv on June 9, 2026 [1]. The model reorders only the top-10 candidate memories selected by ConvMemory v1, boosting the primary evaluation metric without altering which memories are returned [1]. ConvMemory v2 is a fine-tuned cross-encoder based on the ms-marco-MiniLM-L-6-v2 architecture and contains 22,713,601 parameters, measured from the released checkpoint [1][2]. It operates as an opt-in second stage that sits after the lightweight ConvMemory v1 reranker and reorders only v1's protected top-10 candidate set [1][2]. Because v2 does not change which ten memories are returned, Recall@10 and Hit@10 are identical to v1 by construction, not by statistical coincidence [2]. On the LoCoMo conversational memory benchmark, evaluated across five random seeds and 4,955 test rows, ConvMemory v2 raised the FULL Mean Reciprocal Rank from v1's 0.5824 to 0.6560, a paired bootstrap gain of 0.0734 with a 95% confidence interval of [+0.0645, +0.0827] [1][2]. The Hit@1 score improved from 0.4440 to 0.5474 [1][2]. The report positions ConvMemory v2 against a more computationally expensive full-pool cross-encoder reference, mxbai-rerank-large-v1, which achieves an MRR of 0.6688 when applied over the top-500 candidates [1][2]. ConvMemory v2 sits 0.013 below that reference on FULL MRR but exceeds it on two raw-dense-hard slices where v1's protected top-10 has higher recall than mxbai's own top-10 [1][2]. The authors describe this advantage as slice-specific rather than a general dominance claim [2]. A four-arm load-bearing ablation study identified candidate-specific memory text as the mechanism behind ConvMemory v2's performance. Removing, shuffling, or replacing that text caused MRR to collapse below the level of raw dense retrieval [1][2]. The model is characterized as a standard recall-preserving cascade pattern with LoCoMo-specific fine-tuning, an explicit anti-shortcut inference contract, and disciplined load-bearing analysis [2]. This report extends the ConvMemory v1 technical report previously posted at arXiv:2605.28062 [1][2].