Memory Type Varies: Empowering LLM Agents for Long-Term Memory with Diverse Strategies

Yi Wen; Derong Xu; Pengyue Jia; Yichao Wang; Yingyi Zhang; Maolin Wang; Junyi Li; Wenlin Zhang

cs.AIOct 8, 2026

Memory Type Varies: Empowering LLM Agents for Long-Term Memory with Diverse Strategies

Authors: Yi Wen, Derong Xu, Pengyue Jia, Yichao Wang, Yingyi Zhang, Maolin Wang, Junyi Li, Wenlin Zhang, +3 more

Organizations: City University of Hong Kong · Huawei Noah’s Ark Lab

Abstract

The memory capabilities of Large Language Models (LLMs) have garnered increasing attention recently. Despite great success achieved, existing retrieval-based memory approaches typically overlook the differences between memories and employ a unified strategy to process all memories, leading to suboptimal performance. Thus, an intuitive question arises: can we categorize memory into different types and select appropriate strategies? However, given the topic-rich, scenario-complex, and boundary-blurred nature of memory scenarios, achieving precise classification of memories is not easy. To address this challenge, we propose a memory multi-class dataset in this paper, termed TriMEM, which provides precise annotations for memory types across diverse scenarios. Building upon this foundation, we propose a novel memory framework, named MemoType, which can adaptively recognize each memory and query type with the learned router model. With the memory and query routing, MemoType can retrieve the memory with corresponding query types and design tailored retrieval strategies, thereby enhancing the retrieval performance. Moreover, we theoretically prove that any single retrieval strategy is subject to a fundamental upper bound on its expected retrieval precision in multi-class corpora, leading to systematic precision degradation. Extensive experiments on three datasets demonstrate that MemoType consistently outperforms existing methods, achieving up to 16.18% improvement in Recall@1.

Figures & tables

Panel of Figure 1 — Figure 1 : Evidence and Motivation: Semantic Clustering of memory and the MemoType framework.

Figure 2 : Three Memory Types in TriMEM (Left) and the proposed MemoType Framework (Right).

Method	Recall@1	NDCG@1	Recall@3	NDCG@3	Recall@5	NDCG@5	Recall@10	NDCG@10
LongMemEval-S
HyDE ( Gao et al., 2023 )	27.23	27.23	53.19	37.82	65.32	42.84	82.34	49.83
Mill ( Jia et al., 2024 )	38.72	38.72	67.66	50.17	78.94	56.10	90.64	61.34
Query2Doc ( Wang et al., 2023a )	25.96	25.96	50.64	36.43	64.89	41.66	81.06	48.38
SeCom ( Pan et al., 2025 )	45.32	45.32	74.04	53.42	83.19	59.42	91.28	64.66
HippoRAG2 ( Gutiérrez et al., 2025 )	50.64	50.64	80.85	60.59	88.51	66.48	93.53	71.05

Table 1 : Retrieval Performance. All methods are based on Contriever as the retriever.

Method	GPT4Judge	F1	BLEU	Rouge1	Rouge2	RougeL	RougeLsum	BERTScore
LongMemEval-S
HyDE ( Gao et al., 2023 )	42.60	9.96	1.44	10.58	4.67	9.31	9.57	83.02
Mill ( Jia et al., 2024 )	42.20	9.91	1.55	10.57	4.54	9.20	9.45	82.96
Query2Doc ( Wang et al., 2023a )	43.00	9.80	1.47	10.44	4.46	9.21	9.43	82.97
SeCom ( Pan et al., 2025 )	44.80	11.01	1.65	11.66	5.14	10.36	10.58	83.26
HippoRAG2 ( Gutiérrez et al., 2025 )	45.40	10.55	1.65	11.15	5.17	9.98	10.17	83.12

Table 2 : Question Answer Performance. All methods are based on Contriever as the retriever.

Strategy	LongMemEval-S			LoCoMo			PerLTQA
Strategy	Recall@1	Recall@3	Recall@5	Recall@1	Recall@3	Recall@5	Recall@1	Recall@3	Recall@5
Key Expansion	51.77	76.62	85.59	25.93	43.25	51.06	66.10	83.64	88.05
Hypothetical Query	53.91	78.29	87.06	24.42	43.00	50.10	62.65	82.90	89.08
Hypothetical Memory	52.03	77.45	86.22	25.23	42.30	51.11	66.24	83.88	88.34
Type-Aware Strategy	59.15	83.19	90.21	31.07	47.73	54.08	71.92	88.05	92.78
Improvement	5.24	4.90	3.15	5.14	4.48	2.97	5.68	4.17	3.70

Table 3 : The ablation study of the proposed type-aware strategy with single retrieval strategy on three datasets.

Figure 3 : The ablation study of the memory pruning module on two benchmarks.

Appendix figures & tables34 assets

Supplementary material from the paper’s appendix.

Appendix

Memory Type	Content Type	Example
EM	Event Occurred	“Sarah and John had dinner at a sushi restaurant downtown.”
EM	Events Planned to Occur	“Next Month, my colleagues are planning to see a beautiful sunset.”
PM	User Information	“Jack’s degree is in Mathematics. ”
	User Preferences and Habits	“I love Italian food, especially pasta and pizza.”
	User Intent	“I want to buy a pair of running shoes”
GM	General Knowledge	“Water boils at 100 degrees Celsius at sea level.”

Appendix

Table 4 : The Memory Classification Criterion in the TriMEM Benchmark. “EM” denotes Episodic Memory, “PM” denotes the Personal Semantic Memory, “GM” denotes the General Semantic Memory.

Figure 4 : The Statistical Information of TriMEM 3 3 3 The percentage of the memory types is a relative ratio, as there is overlap between different memory types. .

Figure 7 : The Contriever Similarity Distribution between TriMem Context with the raw context on two benchmarks.

Type	Model	TriMEM(In Distribution)				PerLTQA(Out of Distribution)
Type	Model	Overall	EM	PM	GM	Profile	Relationship	Events
ACC (100 %)
Prompting	Qwen3-8b	71.90	81.65	71.37	62.68	88.65	57.21	46.23
Prompting	Gemini-2.5-Flash	77.84	82.47	76.72	74.33	100.00	75.06	91.83
Prompting	GPT-4o-Mini	83.55	85.05	87.18	78.42	100.00	89.47	92.00
Prompting	Qwen3-32b	77.75	87.82	80.88	64.55	100.00	95.64	87.55

Appendix

Table 5 : The classification Performance on TriMEM with four LLM models. “EM” denotes Episodic Memory, “PM” denotes the Personal Semantic Memory, “GM” denotes the General Semantic Memory.

Panel of Figure 8 — Figure 8 : The Distribution of three Similarity Metrics (Word Overlap Rate, BLEU Score, and 3-gram Overlap Rate) between Transferred Memory and Original Memory on LongMemEval-S and LoCoMo datsets. The red dotted line indicates a similarity threshold of 0.5. The majority of samples fall in the low similarity region, indicating a significant difference between the transformed text and the original text.

Panel of Figure 10 — Figure 10 : The Frequency Distributions of three Similarity Metrics between Transferred Memory and Original Memory on LongMemEval-S dataset. The red/green dashed lines indicate the mean and median, respectively. It is evident that all three indicators are concentrated in the lower range.

Dataset	#Inst.	#QA	#Sess.	Sess./Inst.	Turns/Sess.	Words/Inst.	#Q-Types
LoCoMo	10	1,986	272	27.2	21.6	15.9 k	5
LongMemEval-S	500	500	25,097	50.2	0 9.8	79.0 k	6
LongMemEval-M	500	500	249,525	499.1	0 9.8	777.6 k	6
PerLTQA	32	2,838	128	4.0	21.7	0 8.9 k	4

Appendix

Table 6 : Statistics of the long-term memory QA datasets used in our study. “#Inst.” counts the top-level units. “Sess./Inst.” and “Turns/Sess.” are averages; “Words/Inst.” is the average number of whitespace-tokenized words per instance.

Method	LongMemEval-S			LoCoMo			PerLTQA
Method	Recall@1	Recall@3	Recall@5	Recall@1	Recall@3	Recall@5	Recall@1	Recall@3	Recall@5
Base Retriever: MPNet
HyDE	25.11	53.19	66.60	27.29	44.76	53.52	30.62	65.68	74.17
Mill	27.45	53.19	64.89	21.00	36.81	44.51	36.15	67.27	74.77
Query2Doc	27.45	54.68	68.72	26.38	46.12	54.03	33.51	65.82	74.63
SeCom	37.66	62.13	73.19	21.45	34.24	41.49	36.19	65.72	73.75

Appendix

Table 7 : The ablation study of different retrievers.

Method	Recall@1	NDCG@1	Recall@3	NDCG@3	Recall@5	NDCG@5	Recall@10	NDCG@10
LongMemEval-S
w/o Classify	43.40	43.40	73.62	54.00	83.83	60.33	94.26	65.79
w/ Classify	50.21	50.21	79.79	60.32	86.81	65.47	92.34	70.15
LoCoMo
w/o Classify	20.75	20.75	36.56	30.86	44.06	34.16	53.93	37.31
w/ Classify	25.13	25.13	41.99	35.88	49.70	39.06	59.47	42.36

Appendix

Table 8 : The ablation study of the proposed classification strategy on four datasets. “w/ Classify” represents adopting the classification strategy to filter memory, “w/o Classify” denotes the opposite.

Strategy	GPT4Judge	F1	BLEU	Rouge1	Rouge2	RougeL	RougeLsum	BERTScore
LongMemEval-S
w/o Prun	55.60	19.55	4.48	21.32	10.17	19.54	19.70	85.03
w/ Prun	56.60	20.31	4.91	22.18	10.70	20.28	20.32	85.22
LoCoMo
w/o Prun	48.39	20.30	4.11	21.36	10.63	19.93	19.91	85.55
w/ Prun	49.19	22.23	4.64	23.28	11.83	21.83	21.86	85.80

Appendix

Table 9 : The ablation study of memory pruning strategy on three datasets. “w/ Prun” denotes adopting the proposed memory pruning strategy, and“w/o Prun” denotes the contrary.

Figure 11 : Strategy Ablation on Retrieval Recall@1 across Three Benchmarks.

Figure 12 : Strategy Ablation on Retrieval Recall@3 across Three Benchmarks.

Method	Recall@1	NDCG@1	Recall@3	NDCG@3	Recall@5	NDCG@5	Recall@10	NDCG@10
LongMemEval-S
BERT	55.74	55.74	81.06	63.40	88.72	68.64	93.62	72.54
DeBERTa	55.32	55.32	81.91	63.19	89.36	68.90	95.74	72.62
RoBERTa	58.09	58.09	83.83	65.27	90.21	70.70	95.53	74.41
Qwen3-1.7B	59.15	59.15	83.19	66.74	90.21	71.60	92.77	74.85
LoCoMo

Appendix

Table 10 : The Ablation Study of BERT Encoder.

Dataset	Method	Calls	Calls/sample	Prompt (M)	Compl. (M)	Total (M)	USD
LoCoMo	A-Mem	11,764	1,176.4	5.73	1.82	7.56	$1.95
	HippoRAG2	11,760	1,176.0	3.97	0.33	4.30	$0.79
	MemoType	1,756	175.6	0.29	0.08	0.37	$0.09
LongMemEval-S	A-Mem	493,574	987.1	326.67	76.50	403.18	$94.90
	HippoRAG2	492,146	984.3	254.34	29.21	283.56	$55.68
	MemoType	19,897	39.8	3.78	1.43	5.21	$1.42

Appendix

Table 11 : The cost analysis across four benchmarks. “Calls” is the number of paid LLM calls during memory ingestion. Due to the time limits, † HippoRAG2 on LongMemEval-M is linearly projected from

5/500

samples.

Figure 13 : The Semantic Clustering Evidence with Contriever Embedding on two Benchmarks.

Panel of Figure 14 — Figure 14 : Transfer Prompt for User-Assistant Conversation

Panel of Figure 15 — Figure 15 : Transfer Prompt for User-User Conversation

Panel of Figure 16 — Figure 16 : Episodic Memory Classification Prompt

Panel of Figure 17 — Figure 17 : Personal Semantic Memory Classification Prompt

Panel of Figure 18 — Figure 18 : General Semantic Memory Classification Prompt

Panel of Figure 19 — Figure 19 : Memory Multi-Classification Prompt

Panel of Figure 20 — Figure 20 : HyDE Prompt

Panel of Figure 21 — Figure 21 : MILL Prompt

Panel of Figure 22 — Figure 22 : Query2Doc Prompt

Panel of Figure 23 — Figure 23 : Answer Generation Prompt for User-Assistant Conversation

Panel of Figure 24 — Figure 24 : Answer Generation Prompt for User-User Conversation

Panel of Figure 25 — Figure 25 : Query Event Extract Prompt

Panel of Figure 26 — Figure 26 : Conversation Event Extract Prompt

Panel of Figure 27 — Figure 27 : Conversation Keywords Expansion Prompt

Panel of Figure 28 — Figure 28 : Memory Pruning Prompt

Panel of Figure 29 — Figure 29 : Response Prompt, which follows [ Lu et al., 2023 , Pan et al., 2025 ]

Panel of Figure 30 — Figure 30 : GPT Judge Prompt, which follows [ Wu et al., 2025 , Zheng et al., 2023 ]

Explore similar work

MemSifter: Offloading LLM Memory Retrieval via Outcome-Driven Proxy Reasoning
Mar 3, 2026Jiejun Tan, Zhicheng Dou, Liancheng Zhang +3Memory-Augmented Language Models Persistent Memory for Language Models
Gaoling School of Artificial Intelligence · Renmin University of China · Beijing, China
LazyMem: Retrieve Broadly, Construct Selectively for Efficient Long-Term Agent Memory
Jul 17, 2026Jing Yu, Yibo Zhao, Jiaming Zhang +1Persistent Memory for Language Models Long-Term Conversational Memory
School of Data Science and Engineering, East China Normal University
HiGMem: A Hierarchical and LLM-Guided Memory System for Long-Term Conversational Agents
Apr 20, 2026Shuqi Cao, Jingyi He, Fei TanPersistent Memory for Language Models Hierarchical Memory
East China Normal University, Shanghai, China · Shanghai Jiao Tong University, Shanghai, China

Mar 3, 2026, cs.IRJ/K move · Enter open · S save

MemSifter: Offloading LLM Memory Retrieval via Outcome-Driven Proxy Reasoning

Jiejun Tan, Zhicheng Dou, Liancheng Zhang, Yuyang Hu+2

Gaoling School of Artificial Intelligence · Renmin University of China · Beijing, China

Memory-Augmented Language Models Persistent Memory for Language Models LLM Memory

Open paper AskPDF arXiv

As Large Language Models (LLMs) are increasingly used for long-duration tasks, maintaining effective long-term memory has become a critical challenge. Current methods often face a trade-off between cost and accuracy. Simple storage methods often fail to retrieve relevant information, while complex indexing methods (such as memory graphs) require heavy computation and can cause information loss. Furthermore, relying on the working LLM to process all memories is computationally expensive and slow. To address these limitations, we propose MemSifter, a novel framework that offloads the memory retrieval process to a small-scale proxy model. Instead of increasing the burden on the primary working LLM, MemSifter uses a smaller model to reason about the task before retrieving the necessary information. This approach requires no heavy computation during the indexing phase and adds minimal overhead during inference. To optimize the proxy model, we introduce a memory-specific Reinforcement Learning (RL) training paradigm. We design a task-outcome-oriented reward based on the working LLM's actual performance in completing the task. The reward measures the actual contribution of retrieved memories by mutiple interactions with the working LLM, and discriminates retrieved rankings by stepped decreasing contributions. Additionally, we employ training techniques such as Curriculum Learning and Model Merging to improve performance. We evaluated MemSifter on eight LLM memory benchmarks, including Deep Research tasks. The results demonstrate that our method meets or exceeds the performance of existing state-of-the-art approaches in both retrieval accuracy and final task completion. MemSifter offers an efficient and scalable solution for long-term LLM memory. We have open-sourced the model weights, code, and training data to support further research.

Abstract

Explore similar work

MemSifter: Offloading LLM Memory Retrieval via Outcome-Driven Proxy Reasoning

LazyMem: Retrieve Broadly, Construct Selectively for Efficient Long-Term Agent Memory

HiGMem: A Hierarchical and LLM-Guided Memory System for Long-Term Conversational Agents