AuthorsRana Salama, Jason Cai, Michelle Yuan et al.
TL;DR
MemInsight uses autonomous attribute based memory augmentation with priority scoring to reach 60.5% Recall@5 on LoCoMo, a 34.0-point gain over DPR.
Read our summary here, or open the publisher PDF on the next tab.
THE PROBLEM
As interactions accumulate, raw historical data grows rapidly and becomes noisy and imprecise, hindering retrieval and degrading agent performance.
This breakdown in memory retrieval harms long term reasoning and personalization, especially for complex tasks like conversational recommendation, question answering, and event summarization.
HOW IT WORKS
MemInsight’s core mechanism combines Attribute Mining, Annotation and Attribute Prioritization, and Memory Retrieval to turn raw dialogues into structured, queryable memory instances.
You can think of MemInsight as giving the agent a card catalog for its memory, where attributes are index cards and priority decides which drawers to open first.
This structured augmentation lets MemInsight retrieve semantically targeted slices of history that a plain context window or vanilla dense RAG cannot isolate or rank effectively.
DIAGRAM
This diagram shows how MemInsight uses refined retrieval to answer a question by augmenting the query and filtering augmented memory.
DIAGRAM
This diagram shows how MemInsight is evaluated on LLM-REDIAL and LoCoMo from augmentation through retrieval to task metrics.
PROCESS
Attribute Mining
MemInsight applies Attribute Mining to each new interaction, extracting entity centric and conversation centric attributes guided by perspective and granularity.
Attribute Granularity
MemInsight chooses turn level or session level Attribute Granularity, generating complementary annotations like events, emotions, and intents over dialogue spans.
Annotation and Attribute Prioritization
MemInsight runs Annotation and Attribute Prioritization to attach attribute value pairs to memory instances and sort them into Basic or Priority augmentation.
Memory Retrieval
MemInsight uses Memory Retrieval with attribute based or embedding based retrieval to fetch top k augmented memories and integrate them into current context.
KEY CONTRIBUTIONS
Structured Autonomous Memory Augmentation
MemInsight introduces Attribute Mining and Annotation and Attribute Prioritization to adapt memory representations while preserving context across extended conversations and tasks.
Augmentation Aware Memory Retrieval
MemInsight designs Memory Retrieval methods, including attribute based retrieval and embedding based retrieval, to filter irrelevant memory while retaining key historical insights.
Multi Task Empirical Validation
MemInsight achieves 60.5% Recall@5 on LoCoMo versus 26.5% for DPR and improves persuasiveness by up to 14% on LLM REDIAL conversational recommendation.
RESULTS
RECALL@k=5
60.5%
+34.0 over RAG Baseline (DPR)
F1 Overall
30.1%
+1.4 over RAG Baseline (DPR)
Genre Match R@1
0.400
+0.080 over Baseline (Claude-3-Sonnet)
Avg. Attributes
7.39
per movie in LLM REDIAL memory augmentation
On the LoCoMo question answering benchmark, which tests single hop, multi hop, temporal, open domain, and adversarial questions, MemInsight’s 60.5% Recall@5 versus 26.5% for DPR shows that attribute guided augmentation greatly improves retrieval. On LLM REDIAL conversational recommendation, MemInsight’s 0.400 genre match Recall@1 versus 0.320 for the Claude-3-Sonnet baseline demonstrates stronger alignment between recommended and ground truth genres.
BENCHMARK
On the LoCoMo question answering benchmark, which tests single hop, multi hop, temporal, open domain, and adversarial questions, MemInsight’s 60.5% Recall@5 versus 26.5% for DPR shows that attribute guided augmentation greatly improves retrieval. On LLM REDIAL conversational recommendation, MemInsight’s 0.400 genre match Recall@1 versus 0.320 for the Claude-3-Sonnet baseline demonstrates stronger alignment between recommended and ground truth genres.
BENCHMARK
RECALL@k=5 on LoCoMo question answering for embedding based retrieval methods.
KEY INSIGHT
MemInsight retrieves only 10 memory items with embedding based retrieval yet reaches 0.400 genre match Recall@1, beating a 144 item baseline at 0.320.
This is surprising because many assume more retrieved context always helps, but MemInsight shows that fewer, attribute filtered memories can be both leaner and more effective.
WHY IT MATTERS
MemInsight unlocks scalable, semantically structured long term memory where agents can autonomously mine, prioritize, and retrieve attributes across sessions.
Builders can now plug MemInsight into LLM agents to get attribute aware RAG that supports multi hop reasoning, persuasive recommendation, and event summarization without hand crafted schemas.
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.
Yijie Zhong, Yunfan Gao, Haofen Wang
· 2026
HingeMem combines Boundary Guided Long-Term Memory, Dialogue Boundary Extraction, Memory Construction, Query Adaptive Retrieval, Hyperedge Rerank, and Adaptive Stop to segment dialogues into element-indexed hyperedges and plan query-specific retrieval. On LOCOMO, HingeMem achieves 63.9 overall F1 and 75.1 LLM-as-a-Judge score, surpassing the best baseline Zep (56.9 F1) by 7.0 F1 without using category-specific QA formats.
Answers use this explainer on Memory Papers.
Checking…