AuthorsTaiheng Pan
TL;DR
ConvMemory v2 uses a recall-preserving top-10 cross-encoder reranker to boost LoCoMo FULL MRR from 0.5824 to 0.6560 while staying ~68× cheaper than full-pool mxbai.
Read our summary here, or open the publisher PDF on the next tab.
THE PROBLEM
ConvMemory v1 already improves over raw dense retrieval but stalls at FULL MRR 0.5824 with Recall@10 0.7798 on LoCoMo.
This means conversational agents leave substantial ranking headroom unused, and naive full-pool cross-encoders like mxbai-rerank-large-v1 are too expensive for practical RAG systems.
HOW IT WORKS
ConvMemory v2 builds a recall-preserving cascade on top of ConvMemory v1, using a fine-tuned ms-marco-MiniLM-L-6-v2 cross-encoder to rescore v1’s protected top-10.
You can think of ConvMemory v2 like a CPU cache over RAM: ConvMemory v1 cheaply fills a high-recall top-10, and ConvMemory v2 spends expensive cross-encoder compute only on this small prefix.
This design lets ConvMemory v2 exploit candidate-specific memory text while guaranteeing Recall@10 and Hit@10 stay identical to ConvMemory v1, something a plain context window or full-pool reranker cannot ensure.
DIAGRAM
This diagram shows how ConvMemory v2 takes ConvMemory v1’s protected top-10, scores query–memory pairs with a cross-encoder, and returns a recall-preserving reordered list.
DIAGRAM
This diagram shows how ConvMemory v2’s ablation study corrupts candidate-specific memory text to test whether token evidence is load-bearing.
PROCESS
Recall-preserving cascade
ConvMemory v2 starts from ConvMemory v1, which builds a dense MPNet top-500 pool and produces a high-recall protected top-10 candidate set.
Inference prompt format
ConvMemory v2 formats each query and memory using QUERY TIME and MEMORY TIME prefixes, attaching candidate positions as lightweight temporal metadata.
Inference procedure
ConvMemory v2 scores exactly ten query–memory pairs with the ms-marco-MiniLM-L-6-v2 cross-encoder and sorts the protected top-10 by these scores.
Recall-preserving merge
ConvMemory v2 appends ConvMemory v1’s unchanged tail from rank 11 onward, guaranteeing Recall@10 and Hit@10 remain identical to ConvMemory v1.
KEY CONTRIBUTIONS
ConvMemory v2 evidence reranker
ConvMemory v2 introduces a recall-preserving top-10 evidence reranker that boosts FULL MRR from 0.5824 to 0.6560 at only 1.71× ConvMemory v1 latency and about 68× lower cost than mxbai top500.
Token-evidence ablation
ConvMemory v2 includes a four-arm token-evidence ablation where corrupting candidate-specific memory text collapses FULL MRR to 0.2506–0.2966, below raw dense retrieval’s 0.3254.
Anti-shortcut inference contract
ConvMemory v2 enforces an anti-shortcut inference contract that rejects gold, teacher, and GPT-derived fields and uses conversation-level splits, with tests like test_evidence_reranker_rejects_forbidden_fields.
RESULTS
FULL MRR
0.6560
+0.0736 over ConvMemory v1 (0.5824)
H@1
0.5474
+0.1034 over ConvMemory v1 (0.4440)
Recall@10
0.7798
Equal to ConvMemory v1 by construction
Latency ms per query
28.646
1.71× ConvMemory v1 and 68.43× cheaper than mxbai top500
These metrics are on the LoCoMo conversational memory benchmark with 4955 pooled test rows. The results show ConvMemory v2 recovers most of the MRR and H@1 headroom above ConvMemory v1 while preserving Recall@10 and staying far cheaper than mxbai-rerank-large-v1 over the top-500.
BENCHMARK
These metrics are on the LoCoMo conversational memory benchmark with 4955 pooled test rows. The results show ConvMemory v2 recovers most of the MRR and H@1 headroom above ConvMemory v1 while preserving Recall@10 and staying far cheaper than mxbai-rerank-large-v1 over the top-500.
BENCHMARK
FULL MRR on LoCoMo conversational memory retrieval.
BENCHMARK
FULL MRR for ConvMemory v2 full text vs ablation variants and raw dense retrieval.
KEY INSIGHT
When ConvMemory v2’s candidate-specific memory text is removed, shuffled, or replaced, FULL MRR collapses to 0.2506–0.2966, below raw dense retrieval’s 0.3254.
This is surprising because many rerankers degrade toward the dense baseline when corrupted, but ConvMemory v2 becomes confidently wrong, proving that aligned token evidence is truly load-bearing.
WHY IT MATTERS
ConvMemory v2 shows that a recall-preserving top-10 cross-encoder can nearly match a strong full-pool reranker while remaining about 68× cheaper on LoCoMo.
This lets builders add high-precision conversational memory retrieval to RAG systems with strict latency budgets, without sacrificing Recall@10 or relying on opaque shortcut features.
Xingyu Lyu, Jianfeng He et al.
· 2026
ADAM combines Anchor extraction, Distribution estimation, Anchor selection, and Query generation to adaptively probe agent memory via an auxiliary generator and entropy based selection. On the EHRAgent benchmark with Llama2-7b-chat, ADAM reaches EQ=77 and ASR=1.00, compared to MEXTRA’s EQ=44 and ASR=0.89.
Okan Bursa
· 2026
Adaptive RAG Memory (ARM) augments a standard retriever–generator stack with a Dynamic Embedding Layer and Remembrance Engine that track usage statistics and apply selective remembrance and decay to embeddings. On a lightweight retrieval benchmark, ARM achieves NDCG@5 ≈ 0.9401 and Recall@5 = 1.000 with 22M parameters, matching larger baselines like gte-small while providing the best efficiency among ultra-efficient models.
Zhanghao Hu, Qinglin Zhu et al.
· 2026
xMemory organises agent memory via segments, memory components, groups, and dynamic maintenance with a sparsity plus semantic faithfulness objective. On LoCoMo with GPT-5 nano, xMemory reaches 38.71 BLEU and 50.00 F1 versus Nemori’s 36.65 BLEU and 48.17 F1, while cutting tokens per query from 9154.76 to 6581.20.
Answers use this explainer on Memory Papers.
Checking…