Learning What to Share and What to Personalize: Hierarchical Strategy Co-Evolution for Agent Memory

Yupeng Han; Shuochen Liu; Kai Zhang; Ze Liu; Zhihong Pan; Xianquan Wang

cs.AIAug 26, 2026

Learning What to Share and What to Personalize: Hierarchical Strategy Co-Evolution for Agent Memory

Authors: Yupeng Han, Shuochen Liu, Kai Zhang, Ze Liu, Zhihong Pan, Xianquan Wang

Organizations: State Key Laboratory of Cognitive Intelligence, University of Science and Technology of China

Abstract

Memory-augmented agents maintain compact user profiles throughout extended conversations, enabling personalized and consistent responses without the need to process the entire dialogue history. The quality of these user profiles relies on the underlying memory management strategy: at each step, the agent must determine what to retain, compress, or discard. However, existing methods typically employ a static, one-size-fits-all strategy established before training. In practice, the optimal memory decision is inherently user-specific and dynamically evolves alongside policy optimization. To address this, we propose HiPS (Hierarchical Personalized Strategy), a framework that decouples memory management into a globally shared foundation and a user-specific adaptive tier. Specifically, HiPS employs Universal Strategy to extract shared principles from cross-persona trajectories, alongside Persona Delta Distillation to generate tailored rules for users whose behaviors diverge from general patterns. Cross-Level Rule Flow dynamically calibrates their boundary by promoting broadly validated personal rules and demoting contradicted global ones. The architecture establishes a co-evolution loop where a mechanism guarantees that all strategy refinements are anchored to task outcomes. Extensive experiments demonstrate consistent improvements over memory-augmented baselines.

Figures & tables

Figure 1: Same conversation, different users, different needs. A shared strategy (left) applies identical rules to all users, losing Maria’s emotional context while over-preserving Alex’s trivial details. A hypothetical per-user independent approach (center) would redundantly rediscover universal rules and fail on cold-start users. HiPS (right) decomposes the strategy into a shared level and per-user adaptive level.

Figure 2: HiPS framework overview. Universal Strategy Distillation (USD) evolves shared strategies $S_{u}$ from cross-persona trajectories. Persona Delta Distillation (PDD) evolves per-user adaptations $\Delta_{p}$ for users whose divergence exceeds $\theta_{\text{div}}$ . Cross-Level Flow promotes widely validated deltas to $S_{u}$ . The active strategy $S_{p}=S_{u}\cup\Delta_{p}$ is injected into the agent prompt via submodular selection.

Method	PersonaMem		PrefEval		PersonaBench				PERMA
Method	32K	128K	Expl.	Impl.	0	0.3	0.5	0.7	C-S	C-M	N-S	N-M
Long Context	42.17	20.74	32.40	25.90	27.77	22.38	19.45	12.59	24.11	22.88	24.82	22.81
RAG	52.41	38.02	47.40	32.40	32.21	29.13	25.28	23.36	51.63	36.84	51.77	35.09
Mem0	48.53	39.67	57.60	46.40	17.43	18.95	19.31	16.49	52.76	51.67	51.49	51.15
A-Mem	55.42	39.88	62.30	52.80	30.09	28.52	25.81	24.10	53.05	53.73	52.91	49.87
LightMem	52.41	36.74	64.20	54.80	19.14	17.83	19.61	17.51	59.43	55.53	59.43	55.53

Table 1: Overall comparison across twelve evaluation settings. PersonaMem (32K/128K) is in-domain; PrefEval (Explicit/Implicit), PersonaBench (4 noise levels), and PERMA (C-S/C-M/N-S/N-M denote Clean/Noise

\times

Single/Multi-Domain) are out-of-domain. Higher is better. Best results are in bold .

	PersonaMem		PERMA
Configuration	32K	128K	C-S	C-M	N-S	N-M
HiPS (Full)	73.49	62.01	66.95	56.56	63.83	56.30
w/o Flow	71.08	60.49	45.39	40.87	43.12	36.50
w/o Gate	69.28	59.63	51.63	44.99	52.62	46.79
w/o USD	66.27	53.70	61.28	51.93	60.14	49.10
w/o PG	67.47	52.22	60.70	51.41	59.43	53.98

Table 2: Ablation results on PersonaMem and PERMA. Cell color intensity indicates drop from Full (

<

3 to

\geq

15 pp). Rows ordered by average drop.

Figure 3: Evolution of

S_{u}

evidence distribution and accuracy throughout training on PersonaMem.

Figure 4: Per-persona impact of persona-specific strategies ( $\Delta_{p}$ ). We compare Full HiPS (with $\Delta_{p}$ ) against w/o PDD (without $\Delta_{p}$ ) on PersonaMem 128K. PDD yields an average improvement of +6.0%, benefiting 20 personas. However, the gains are highly heterogeneous: P15 (finance enthusiast) gains +18.6%, while P6 (marketing specialist) suffers -3.3%. We highlight four representative personas to illustrate why.

Method	Qwen2.5-7B Instruct	GPT-4o -mini	Gemini 2.5 flash	GPT-5
RAG	52.41	51.58	60.83	63.74
A-Mem	55.42	55.91	62.74	65.93
$\blacktriangledown$ Optimized w/ Qwen2.5-7B-Instruct
HiPS ( $S_{u}$ only)	57.83	56.74	64.25	67.18
HiPS ( $S_{u}$ + $\Delta_{p}$ )	61.24	59.83	67.45	70.16
$\blacktriangledown$ Optimized w/ GPT-4o-mini

Table 3: Cross-model transferability of strategies (without RL policy). Strategies are distilled with one LLM and injected into others’ prompts for inference.

Figure 5: Scaling analysis on PersonaMem (128K). We increase the total dialogue tokens (each evolution round processes 4K tokens) and report the resulting memory bank size (left) and memory evolving time (right).

Appendix figures & tables10 assets

Supplementary material from the paper’s appendix.

Appendix

Statistic	PersonaMem
Tokens per history	~32k	~128k	~1M
# QA pairs	589	2727	2674
# Sessions per history	10	20	60
Avg. # utterances	167.1	758.3	3607.9

Appendix

Table 4: Statistics of the PersonaMem dataset at different context lengths. Token counts denote the approximate total context length per interaction history; utterance counts are averaged over histories.

Statistic	Value
Explicit queries	1,000
Implicit queries	1,000
Maximum inserted conversations	24
Maximum inserted turns	326
Avg. turns / conversation	13.58
Total tokens	108,102

Appendix

Table 5: Dataset statistics for PrefEval multiple-choice classification.

User	Queries	Corpus	Conv.	AI	E-com.
1	48	110	84	23	3
2	43	90	78	8	4
3	42	64	51	12	1
4	46	85	71	14	0
5	44	84	59	21	4
6	40	94	79	14	1

Appendix

Table 6: Statistics of the PersonaBench subset across six users. Corpus is the sum of Conv. , AI , and E-com. .

Metric	P1	P2	P3	P4	P5	P6	P7	P8	P9	P10	Total
User Demographics
Age	35-44	25-34	65+	55-64	35-44	18-24	25-34	35-44	25-34	25-34	–
Gender	M	M	F	M	M	M	M	M	F	M	–
Education	Grad.	Univ.	–	Univ.	–	Grad.	Univ.	–	Grad.	Univ.	–
Country	CA	MX	FI	US	AU	UK	CH	IL	RU	BE	–
Interaction Statistics (Base Dataset: Clean / Noisy)

Appendix

Table 7: Comprehensive statistics of the dataset across multiple dimensions. The table summarizes user demographics for profiles, followed by statistics for the base (comparing Clean and Noise) and the style-aligned long-context dataset. Note: For User Demographics Profile (P), Gender: M (Male), F (Female); Education: Grad. (Graduate Degree), Univ. (University); Country codes follow ISO 3166-1 alpha-2. In Interaction Statistics, # denotes the count, and (k) represents thousands of tokens.

Method	Recall facts	Suggest ideas	Latest prefs	Prefs evolve	Update reasons	Aligned recs	New Scenarios	Overall
$\blacktriangledown$ 32K memory corpus data
Long Context	41.38	22.22	-	42.22	72.41	33.33	26.67	42.17
RAG	37.93	22.22	-	64.44	86.21	60.00	33.33	52.41
Mem0	47.93	19.41	-	46.61	79.58	57.04	42.75	48.53
A-Mem	45.71	18.52	-	73.33	79.31	66.67	33.33	55.42
LightMem	47.06	12.90	-	66.19	75.76	38.18	24.56	52.41

Appendix

Table 8: Category-wise accuracy (%) on PersonaMem under 32K and 128K interaction histories. “–” indicates the category is not available in the dataset.

Method	Travel	Entertain	Lifestyle	Shop	Education	Professional	Pet	Overall
$\blacktriangledown$ Explicit Memory
Long Context	33.90	42.30	37.80	24.10	24.70	23.70	8.00	32.40
RAG	48.50	55.90	52.40	40.30	40.60	38.00	25.90	47.40
Mem0	60.10	64.00	62.30	49.70	53.20	50.20	38.70	57.60
A-Mem	63.90	69.20	65.60	56.70	56.80	55.50	43.30	62.30
LightMem	66.10	70.90	67.20	59.30	57.50	56.30	47.00	64.20

Appendix

Table 9: Domain-wise accuracy (%) on PrefEval multiple-choice classification under Explicit vs. Implicit preference.

Hyperparameter	Symbol	Value
Universal Strategy Distillation (USD)
USD distillation frequency	$K_{1}$	20
Distillation trajectories per cycle	$k$	5
Trajectory buffer size	—	50
Promotion threshold	$\theta_{\text{val}}$	0.1
Demotion threshold	$\theta_{\text{rev}}$	0.02

Appendix

Table 10: Hyperparameter settings used for HiPS strategy co-evolution.

Figure 6: Parameter sensitivity analysis of HiPS on the PersonaMem 128K benchmark. The red stars indicate the selected default values. From top-left to bottom-right, the subplots demonstrate the impact on prediction accuracy of the USD frequency

K_{1}

, PDD frequency

K_{2}

, strategy token budget

B

, divergence threshold

\theta_{\text{div}}

, flow threshold

\theta_{\text{flow}}

, validation threshold

\theta_{\text{val}}

, revision threshold

\theta_{\text{rev}}

, and quality weight

\lambda

Panel of Figure 7 — Figure 7: The complete prompt template used for Universal Strategy Distillation.

Panel of Figure 8 — Figure 8: The complete prompt template used for Persona Delta Distillation.

Agent Memory Management

Agent memory management studies how AI agents store, update, retrieve, compress, and forget information across long-running tasks.

See all 446 papers

Hierarchical Memory Orchestration for Personalized Persistent Agents
Apr 2, 2026Junming Liu, Yifei Sun, Weihua Cheng +4LLM Agent Memory Persistent Memory for Language Models
Shanghai Artificial Intelligence Laboratory, Shanghai, China · The City University of New York, New York, USA
Controllable Memory Usage: Balancing Anchoring and Innovation in Long-Term Human-Agent Interaction
Jan 8, 2026Muzhao Tian, Zisu Huang, Xiaohua Wang +8Long-Term Conversational Memory Conversational Agents
College of Computer Science and Artificial Intelligence, Fudan University Shanghai Key Laboratory of Intelligent Information Processing
AdaMem: Adaptive User-Centric Memory for Long-Horizon Dialogue Agents
Mar 17, 2026Shannan Yan, Jingchen Ni, Leqi Zheng +6LLM Agent Memory Long-Term Conversational Memory
Tsinghua University · WeChat Vision, Tencent Inc.

Apr 2, 2026J/K move · Enter open · S save

Hierarchical Memory Orchestration for Personalized Persistent Agents

Junming Liu, Yifei Sun, Weihua Cheng, Haodong Lei+3

Shanghai Artificial Intelligence Laboratory, Shanghai, China · The City University of New York, New York, USA

LLM Agent Memory Persistent Memory for Language Models Agent Memory Management Hierarchical Memory

Open paper AskPDF arXiv

While long-term memory is essential for intelligent agents to maintain consistent historical awareness, the accumulation of extensive interaction data often leads to performance bottlenecks. Naive storage expansion increases retrieval noise and computational latency, overwhelming the reasoning capacity of models deployed on constrained personal devices. To address this, we propose Hierarchical Memory Orchestration (HMO), a framework that organizes interaction history into a three-tiered directory driven by user-centric contextual relevance. Our system maintains a compact primary cache, coupling recent and pivotal memories with an evolving user profile to ensure agent reasoning remains aligned with individual behavioral traits. This primary cache is complemented by a high-priority secondary layer, both of which are managed within a global archive of the full interaction history. Crucially, the user persona dictates memory redistribution across this hierarchy, promoting records mapped to long-term patterns toward more active tiers while relegating less relevant information. This targeted orchestration surfaces historical knowledge precisely when needed while maintaining a lean and efficient active search space. Evaluations on multiple benchmarks achieve state-of-the-art performance. Real-world deployments in ecosystems like OpenClaw demonstrate that HMO significantly enhances agent fluidity and personalization.

Abstract

Agent Memory Management

Explore similar work

Hierarchical Memory Orchestration for Personalized Persistent Agents

Controllable Memory Usage: Balancing Anchoring and Innovation in Long-Term Human-Agent Interaction

AdaMem: Adaptive User-Centric Memory for Long-Horizon Dialogue Agents