Paper deep dive
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
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 91%
Last extracted: 8/27/2026, 5:08:50 AM
Summary
The paper introduces HiPS (Hierarchical Personalized Strategy), a framework for memory-augmented agents that decouples memory management into a globally shared universal strategy and a user-specific adaptive tier. HiPS employs Universal Strategy Distillation (USD) to extract shared principles from cross-persona trajectories, Persona Delta Distillation (PDD) to generate tailored rules for divergent users, and Cross-Level Rule Flow to dynamically calibrate the boundary between these tiers. This co-evolutionary approach addresses the limitations of static, one-size-fits-all memory strategies by adapting to individual user behaviors while maintaining efficiency.
Entities (6)
Relation Signals (7)
HiPS → contains → Persona Delta Distillation
confidence 95% · HiPS contains three core mechanisms: ... (2) Persona Delta Distillation (PDD)...
HiPS → contains → Cross-Level Rule Flow
confidence 95% · HiPS contains three core mechanisms: ... (3) Cross-Level Rule Flow...
HiPS → contains → Universal Strategy Distillation
confidence 95% · HiPS contains three core mechanisms: (1) Universal Strategy Distillation (USD)...
Cross-Level Rule Flow → calibrates → boundary between Universal Rule Base and Personal Strategy
confidence 90% · Cross-Level Rule Flow dynamically calibrates the boundary between these tiers by promoting broadly validated personal rules to the universal level and substituting contradicted global rules with targeted PDD revisions.
Universal Strategy Distillation → generates → Universal Rule Base
confidence 90% · USD evolves shared rules using cross-persona trajectories... USD generates structured revisions that validate or refine rules.
Persona Delta Distillation → generates → Personal Strategy
confidence 90% · PDD formulates adaptive, per-user rules specifically for individuals whose behaviors diverge from the population norm.
HiPS → improves → memory-augmented baselines
confidence 85% · Extensive experiments demonstrate consistent improvements over memory-augmented baselines.
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
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 \textbf{HiPS} (\textbf{Hi}erarchical \textbf{P}ersonalized \textbf{S}trategy), a framework that decouples memory management into a globally shared foundation and a user-specific adaptive tier. Specifically, HiPS employs \textbf{Universal Strategy} to extract shared principles from cross-persona trajectories, alongside \textbf{Persona Delta Distillation} to generate tailored rules for users whose behaviors diverge from general patterns. \textbf{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.
Tags
Links
- Source: https://arxiv.org/abs/2608.25329v1
- Canonical: https://arxiv.org/abs/2608.25329v1
Trouble viewing inline? Open PDF directly →
Full Text
69,899 characters extracted from source content.
Expand or collapse full text
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 State Key Laboratory of Cognitive Intelligence, University of Science and Technology of China yupenghan@mail.ustc.edu.cn kkzhang08@ustc.edu.cn Abstract Memory-augmented agents maintain compact user profiles throughout extended conversa- tions, enabling personalized and consistent re- sponses without the need to process the entire dialogue history. The quality of these user pro- files relies on the underlying memory manage- ment 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 be- fore training. In practice, the optimal memory decision is inherently user-specific and dynami- cally evolves alongside policy optimization. To address this, we propose HiPS (Hierarchical Personalized Strategy), a framework that de- couples memory management into a globally shared foundation and a user-specific adap- tive tier. Specifically, HiPS employs Univer- sal 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 contra- dicted global ones. The architecture estab- lishes a co-evolution loop where a mechanism guarantees that all strategy refinements are an- chored to task outcomes. Extensive experi- ments demonstrate consistent improvements over memory-augmented baselines 1 . 1 Introduction Large language models (LLMs) are increasingly deployed as personalized conversational agents, which require maintaining coherent and tailored interactions for individual users (Jiang et al., 2025; Zhao et al., 2025). However, the inherent con- straints of finite context windows prohibit LLMs from retaining unbounded dialogue histories (Xu * Corresponding author. 1 https://github.com/Hyp26cs/HiPS Incoming Chunks : I‘m frustrated with my advisor‘s tech stock picks. Thinking about switching to index funds. Maria lFinancial Advisor lEmotionally engaged with finance Alex lCasual investor, Data-driven lDetached from finance news Shared Only Per-User Only (No Sharing) HiPS (Ours) Shared Rule Base Conflict Resolution Change Tracking Temporal Anchoring Maria Profile: Financial Investor State: Wants to switch from tech stocks to index funds Action: Update portfolio allocation Alex Profile: Financial Investor State: Frustrated with advisor's picks demands index fund migration Action: Migrate capital immediately Policy too rigid! Trust-erosion context dropped Failed to track relationship crisis Policy too broad! Emotional noise leaked in Corrupted data-driven metrics Conflict Resolution Change Tracking Rule: Trust-Erosion Filter Conflict Resolution Change Tracking Rule: Quant Metrics Only Maria’s Rule Base Alex’s Rule Base Token Bloat! Maria Profile: High-trust advisory client State: Client relationship at risk; User frustrated with advisor's picks Action: Pause active tech stock advice Alex Profile: Data-driven investor State: Reallocating portfolio due to tech stock underperformance Action: Execute index fund switch l Severe cold-start overhead for every single new persona l Linear replication of generic rules causes parameter/token bloat Scalability Crisis! Universal Strategy Conflict Resolution Change TrackingTemporal Anchoring Quant Metrics Only Maria’s Personal Strategy Trust-Erosion Filter Alex’s Personal Strategy Quant Metrics Only Profile: Financial Investor Advisor Expert State: Shift to index funds; Recaptured trust crisis Action: Update portfolio; Pause active tech tips Profile: Final investor; Data- driven Retail State: Shift to index funds; Filter To low ROI metrics Metrics:Update portfolio; Execute index switch MariaAlex Emotional nuances fully recaptured Irrelevant financial Chatter filtered 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. et al., 2026b; Gao et al., 2025). Moreover, naive storage and retrieval of raw dialogue fail to cap- ture the dynamic, evolving user preferences (Zhang et al., 2026b). This challenge motivates the inte- gration of external memory systems, requiring the agent to dynamically evaluate at each turn whether to retain, compress, or discard incoming context. Along this line, most existing memory systems employ static workflows, converting raw dialogue into external memory via predefined extraction and compression rules (Chhikara et al., 2025; Xu et al., 2026b; Fang et al., 2026). However, they lack the capacity to learn from interaction feed- back or adapt to diverse user behaviors. Recent approaches formalize memory operations as learn- able actions, optimizing update policies through reinforcement learning (RL) (Ouyang et al., 2026; Yu et al., 2026b; Zhou et al., 2026) or natural- language strategy distillation (Xu et al., 2026a). While more adaptive, they impose a strictly user- agnostic management paradigm. When optimizing for population-averaged rewards, the learning sig- nals for niche behaviors are overwhelmed by the majority (Poddar et al., 2024), resulting in a one- arXiv:2608.25329v1 [cs.AI] 26 Aug 2026 size-fits-all compromise. Moreover, methods utiliz- ing explicit strategies (Xu et al., 2026a) keep these rules frozen during training. Blind to online feed- back, the prescribed strategy misaligns with actual rollout trajectories as the policy evolves, making it imperative that strategies be both personalized and adaptive (Hu et al., 2026b). Yet, individualiz- ing every operation is redundant, as foundational rules are universally beneficial while niche behav- iors require tailored handling. Since the optimal boundary is unknown a priori, this naturally raises the question: Can we dynamically discover and co-evolve this partition using on-policy evidence? To this end, we introduce HiPS (Hierarchical Personalized Strategy), a framework that decom- poses the management strategy into a shared uni- versal level and a per-user adaptive level, dynam- ically learning the boundary between them from on-policy evidence. As illustrated in Fig. 1, HiPS contains three core mechanisms: (1) Universal Strategy Distillation (USD) evolves shared rules using cross-persona trajectories. By contrasting high- and low-reward episodes, USD generates structured revisions that validate or refine rules. (2) Persona Delta Distillation (PDD) formulates adaptive, per-user rules specifically for individu- als whose behaviors diverge from the population norm. A divergence criterion selectively identifies users who need personalization, thereby preventing overfitting and noise for those adequately served by universal rules. Finally, (3) Cross-Level Rule Flow dynamically calibrates the boundary between these tiers by promoting broadly validated personal rules to the universal level and substituting con- tradicted global rules with targeted PDD revisions. Unlike previous approaches that freeze strategies before training (Xu et al., 2026a), HiPS interleaves strategy discovery and policy optimization within a unified, continuous loop. In this paradigm, the active strategy directs model rollouts to generate trajectories that subsequently serve as empirical feedback for further strategy refinement. To pre- vent self-reinforcing confirmation bias (Tan et al., 2025a), strategy evolution is anchored strictly to objective task outcomes, reserving rule adherence exclusively for policy optimization. Our contribu- tions are summarized as below: •We formulate the strategy personalization prob- lem, demonstrating that the boundary between shared and user-specific memory rules must be discovered empirically rather than predefined. •We propose HiPS, an RL framework that co- evolves a universal strategy and user-specific adaptive rules alongside a policy via evidence- based distillation and cross-level rule flow. •Extensive experiments demonstrate consistent performance gains. We uncover a component importance flip where universal rules dominate in-domain tasks and adaptive mechanisms drive out-of-domain generalization. 2 Related Work Memory architectures.LLM agent memory sys- tems manage how historical information is stored, organized, and retrieved within a bounded context window. Memory bank approaches apply segmen- tation, summarization, and selective forgetting to maintain long-term quality (Chhikara et al., 2025; Xu et al., 2026b; Fang et al., 2026). Structured indices such as tree- and graph-based retrieval im- prove access efficiency (Packer et al., 2024; Li et al., 2025). Personalization-oriented systems extract and maintain user profiles for downstream condi- tioning (Jiang et al., 2025). Despite their diversity, these systems share two fundamental limitations: their management strategies are fixed, rendering them incapable of learning from interaction feed- back, and shared, imposing identical rules across all users regardless of behavioral differences. RL-trained and strategy-based memory policies. To overcome the rigidity of fixed pipelines, recent studies formulate memory operations as learnable actions via reinforcement learning (RL) (Yan et al., 2025; Wang et al., 2026; Yu et al., 2026b; Zhou et al., 2026; Ouyang et al., 2026). While these methods effectively learn parametric and procedu- ral policies, they fail to develop declarative and interpretable strategies: management behaviors are implicitly encoded within model parameters, ren- dering them neither inspectable nor editable. A parallel line of research formulates strategies as natural-language rules. For instance, MemCoE (Xu et al., 2026a) optimizes management guidelines using TextGrad before RL training, subsequently injecting them as system prompts for GRPO fine- tuning. EverMemOS (Hu et al., 2026a) intro- duces skill retirement through experience cluster- ing and distillation, while MemSkill (Zhang et al., 2026a) reframes traditional static memory opera- tions into learnable and evolvable “memory skills”, dynamically optimizing both skill selection and Structured Diff 휹 푼푺푫 푽,푯,푹 Validate VRevise R Hypothesize H Low 푹 풂풏풔 Group Persona High 푹 풂풏풔 Group Universal Rule Base S u Supported Conflict Resolution Change Tracking Established Temporal Anchored Supported Compression Tentative Universal Strategy Distillation 2 PG(r,s) From trajectories + R ans PG Computation (Predictive Gain) Strategy Injection & Co-Evolution 5 Agent System Prompt [Universal Guideline] From 푺 풖 [User-Specific Adjustments] From ∆ 풑 GRPO Policy 흅 휽 Update 휽 풎 풕 =흅 휽 (풄 풕 ,풎 풕#ퟏ ;푺 풑 ) Budgeted Rule Selection (Submodular) |퐒|3퐁 퐦퐚퐱 " 퐫∈퐬 퐏퐆퐫.(ퟏ− 퐬;퐫 퐦퐚퐱 훔(퐫,퐬)) 푹 풇풐풍풐풘 (rule compliance) 푹 풂풏풔 (answer correctness) Reward (Signal Separation) Active Strategy 푺 풑 =푺 풖 ⋃∆ 풑 Personal Strategy ∆ 풑 Universal Strategy S u R ans Path: For GRPO optimization, trajectory ranking, and strategy updates.R follow Path : For GRPO optimization only; no feedback to strategy updates. Personal Delta Rule Base ∆ 풑 Trust-Erosion Filter Metric-Only Summary Find-grained Change Tracking TentativeSupportedEstablished Refining Behavior-Conditioned Distillation Observed Trajectory Expected Under 푺 풖 Divergence 푫 풑 Persona Delta Distillation 4 Divergence Gate 푫 풑 ≥휽 풅풆풗 Yes ， go to ∆ 풑 No ， use 푺 풑 only Specification Cross-Level Rule Flow Generalization Learned Boundary θ dev 푺 풖 → ∆ 풑 ∆ 풑 → 푺 풖 3 On-Policy Memory Trajectories (Ranked by 푹 풂풏풔 ) Context 풄 풕 Memory 풎 풕 Answer Reward 푹 풂풏풔 Start 1 Every K 1 StepsEvery K 2 Steps Figure 2: HiPS framework overview. Universal Strategy Distillation (USD) evolves shared strategiesS u from cross- persona trajectories. Persona Delta Distillation (PDD) evolves per-user adaptations∆ p for users whose divergence exceedsθ div . Cross-Level Flow promotes widely validated deltas toS u . The active strategyS p = S u ∪ ∆ p is injected into the agent prompt via submodular selection. refinement. Although these approaches yield inter- pretable strategies, the resulting rules are inherently globally shared and frozen prior to training: a single guideline serves all users and remains static once optimized. HiPS differs by employing user- indexed strategies that dynamically co-evolve with the policy driven by on-policy evidence, rather than relying on static, globally shared rules. 3 Preliminary We consider a conversational setting in which a memory-augmented agent interacts with a userp over multiple sessions. The dialogue history is seg- mented into contextual chunksc 1 ,...,c K , pro- cessed sequentially by the agentπ θ . To facilitate long-term personalization, the agent maintains a compact memory statem t that is updated as new information arrives. The memory update mecha- nism is governed by a set of management rules, denoted asS p , which is injected into the agent’s system prompt to regulate its actions at step t: m t = π θ (c t ,m t−1 ; S p ).(1) Prior work treats the management strategy as a single, globally shared entity that is fixed prior to training. This introduces two primary limitations: the strategy lacks user-specific adaptability, and it remains static as the policy evolves. To overcome these limitations and enable dynamic adaptation, we decomposeS p into shared and personalized components, allowing this partition to co-evolve with the policy during training. Formally, S p = S u ∪ ∆ p ,(2) whereS u comprises universally beneficial rules applicable to the entire population, and∆ p contains adaptive rules tailored to the behavioral patterns of userp. After processing allKdialogue chunks, the agent generates a final response to a queryq conditioned on the terminal memory state m K . 4 Method In this section, we propose HiPS, a framework that decouples the memory management strategy into a shared universal tier and a user-specific adaptive tier. Specifically, Universal Strategy Distillation (USD) (Section 4.1) abstracts broadly applicable principles from cross-persona trajectories. Sub- sequently, Persona Delta Distillation (PDD) (Sec- tion 4.2) captures idiosyncratic rules specifically for users exhibiting divergent behaviors. Finally, a Cross-Level Rule Flow mechanism (Section 4.3) calibrates the partition by migrating rules between the two tiers as evidence accumulates. The overall architecture of HiPS is illustrated in Fig. 2. 4.1 Universal Strategy Distillation (USD) Existing strategies are either manually crafted or learned from data, yet both are subsequently frozen, thereby failing to incorporate dynamic evidence ac- cumulated as the policy evolves. To break this static paradigm,S u should be continuously refined using live trajectories. While contrastive analy- sis of trajectories can identify promising patterns, distilling these insights into reliable rule updates presents two challenges: (1) Conventional free- form feedback treats the guideline as a monolithic text, lacking per-rule granularity. (2) Without gran- ular evidence tracking, a rule proposed in one cycle may be arbitrarily overwritten in the next. USD resolves these issues by utilizing structured diffs to decompose updates into rule-level operations, cou- pled with a rigorous evidence tracking mechanism to maintain a persistent validation history. Contrastive feedback as structured diff.At reg- ular training intervals, USD samples a persona- balanced set of the top-kand bottom-ktrajecto- ries to prevent dominant users from skewing the optimization signal. An LLM meta-optimizer is prompted to compare these contrastive sets. Instead of yielding a free-form summary, the optimizer out- puts a structured diff, a discrete set of operations applied to the current rule set (see Appendix F for prompt). Formally, δ USD =V : [·], H : [·], R : [·],(3) where V signifies validating an existing rule, H represents hypothesizing a new rule, and R de- notes revising or retiring a rule. Crucially, new hypotheses must strictly adhere to the “[Label]: When [condition], [action]” format and prescribe explicit management behaviors. Evidence tracking. Since a single distillation cycle may yield spurious hypotheses, USD incor- porates strict evidence tracking to filter out noise. Each rule’s evidence level, includingsupported, established , andtentative(see Appendix E for details) is iteratively updated based on its val- idation signals: consistent validation operations promote the rule, whereas revision or contradiction operations demote it. To complement the LLM judgment with a data-driven safeguard, rules are au- tomatically promoted if their predictive gain (Eq. 4) exceeds θ val , and demoted if it falls below θ rev . 4.2 Persona Delta Distillation (PDD) Under globally shared strategies, rules benefiting a minority receive diluted signals while locally harm- ful ones persist. Although per-user adaptation re- solves this, naively generating∆ p for everyone wastes computational budget and introduces noise for those adequately served byS u . To mitigate this, PDD employs a divergence criterion to gate personalization, selectively identifying users who genuinely require adaptive rules. Divergence-gated personalization.To quantify behavioral deviation, we first evaluate the extent to which each universal rule correlates with task success. We define the predictive gain of a ruler as the absolute difference in the expected success rate (denoted byY ∈0, 1) between trajectories that comply with r and those that do not: PG(r) = P (Y = 1| r + )− P (Y = 1| r − ) , (4) wherer + andr − indicate rule compliance and non- compliance, respectively. PG provides a proxy for rule importance: rules with high PG strongly predict task outcomes, while near-zero PG signals irrelevance. Crucially, PG is used only for relative ranking and threshold-based gating, not as an ab- solute measure of rule quality. Building upon this, we define the divergence of userpas the aggregate discrepancy in rule efficacy between the specific user and the broader population: D p = X r∈S u |PG(r | p)− PG(r)|.(5) Users exhibiting a divergenceD(p)≥ θ div trigger the PDD module. Conversely, users below this threshold are deemed sufficiently covered byS u and bypass the personalization step. This gating mechanism resolves the cold-start dilemma by de- faulting low-divergence users to the universal base- line, optimizing resource allocation. Behavior-conditioned distillation. For users identified as divergent, PDD prompts an LLM to formulate management behaviors that strictly dif- fer fromS u , using the universal rules as an an- choring context. To prevent∆ p from degenerat- ing into mere factual knowledge, the generated rules are constrained to be management-oriented to prescribe actions rather than asserting facts and behavior-conditioned. This ensures robust general- ization to unseen users at inference time. 4.3 Cross-Level Rule Flow IfS u and∆ p evolve in strict isolation, the bound- ary between them becomes brittle. Without a dy- namic transfer mechanism, a personalized rule that ultimately proves globally beneficial remains per- manently trapped in∆ p , while a universal rule con- tradicted by a subset of users persists erroneously inS u . To resolve this structural rigidity, the Cross- Level Rule Flow dynamically calibrates the par- tition by migrating rules between tiers based on accumulated cross-persona evidence. Generalization (∆ p → S u ). When a∆ p rule at “supported” level or above appears in≥ θ flow frac- tion of personas that have∆ p , it is promoted toS u and pruned from individual deltas. Intuitively, a management pattern emerging across diverse users signifies a universal principle rather than a local- ized adaptation. To ensure robust aggregation, we match rules via semantic similarity, consolidating equivalent behaviors despite syntactic variations. Specialization (S u → ∆ p ). When USD revises or demotes a rule withinS u , users who previously benefited from the original rule risk losing criti- cal guidance. To mitigate this, PDD hypothesizes persona-specific replacements, effectively restor- ing the management behavior in a localized form. This mechanism prevents a one-size-fits-all revi- sion from deteriorating the performance of minor- ity users who relied on the deprecated rule. Together, generalization and specialization es- tablish a cycle. Local discoveries that prove widely beneficial are assimilated intoS u , whereas univer- sal rules that are contradicted by specific subpopu- lations are replaced by tailored adaptations in ∆ p . 4.4 Strategy Injection and Co-Evolution In this co-evolutionary framework, the active strat- egy directs policy rollouts, while the resulting tra- jectories provide empirical feedback for subsequent strategy refinement. We link these phases using budgeted rule injection and a combined reward that decouples strategy evolution from adherence. Submodular selection.The active strategyS p = S u ∪∆ p is constrained by a maximum token budget B. To select the most representative and diverse set of rules, we greedily maximize a submodular coverage objective penalized by redundancy: max S⊆S p P r∈S |r|≤B X r∈S PG(r)· 1− max s̸=r σ(r,s) ,(6) whereσ(·,·)denotes semantic similarity. The to- ken budget is allocated between the universal and personalized levels proportionally to their aggre- gate predictive gains, subject to a minimum floor for each level as detailed in Appendix C. Guideline-aligned reward. To align the policy with the discovered strategies, we define a persona- aware adherence reward: R follow (τ,p) = P r∈S p PG(r)· h(τ,r) P r∈S p PG(r) ,(7) whereh(τ,r)∈ [0, 1]is a compliance check. The combined reward for Group Relative Policy Opti- mization (GRPO) (Shao et al., 2024) isR(τ,p) = R ans (τ ) +λ·R follow (τ,p) . Consequently, the iden- tical trajectory naturally yields varying adherence scores depending on the user’s specific ∆ p . Anti-circularity and co-evolution.A rule could validate itself if compliance inflates the signal used to select which trajectories inform strategy updates. We disrupt this direct path by ranking the distilla- tion buffer exclusively byR ans , while restricting R follow to GRPO advantage computation. An in- direct path remains whereR follow shapes the pol- icy, which alters rollout behavior and subsequently affects futureR ans . However, its indirect path is benign when rule compliance genuinely improves task success, which is precisely the condition under which strategy refinement should occur. A vulner- ability arises only when compliance is rewarded without corresponding task improvement. Sepa- ratingR ans for distillation fromR follow for policy optimization prevents this reinforcement loop. The detailed training procedure, detailed in Appendix F, integrates these components into a unified loop. 5 Experiments 5.1 Experimental Setup Benchmarks. We evaluate on four personalized memory benchmarks: PersonaMem (Jiang et al., Method PersonaMemPrefEvalPersonaBenchPERMA 32K128KExpl.Impl.00.30.50.7C-SC-MN-SN-M Long Context42.17 20.7432.40 25.9027.77 22.38 19.45 12.5924.11 22.88 24.82 22.81 RAG52.41 38.0247.40 32.4032.21 29.13 25.28 23.3651.63 36.84 51.77 35.09 Mem048.53 39.6757.60 46.4017.43 18.95 19.31 16.4952.76 51.67 51.49 51.15 A-Mem55.42 39.8862.30 52.8030.09 28.52 25.81 24.1053.05 53.73 52.91 49.87 LightMem 52.41 36.7464.20 54.8019.14 17.83 19.61 17.5159.43 55.53 59.43 55.53 MemAgent59.93 50.8677.40 64.4030.74 28.34 24.47 23.5353.98 53.90 55.74 48.84 Mem-α58.43 48.1875.70 63.1019.31 18.01 17.35 16.6452.19 51.67 54.46 46.53 MemSkill64.45 55.3782.60 68.4031.19 29.89 24.85 23.3662.83 51.92 60.14 53.21 HiPS (Ours)73.4962.0189.2069.4032.2729.7625.9925.0966.9556.5663.8356.30 Table 1: Overall comparison across twelve evaluation settings. PersonaMem (32K/128K) is in-domain; PrefE- val (Explicit/Implicit), PersonaBench (4 noise levels), and PERMA (C-S/C-M/N-S/N-M denote Clean/Noise× Single/Multi-Domain) are out-of-domain. Higher is better. Best results are in bold. 2025) for preference evolution over varying con- text scales; PrefEval (Zhao et al., 2025) for explicit and implicit queries with 50 distraction turns; Per- sonaBench (Tan et al., 2025b) for personalized retrieval and QA over corpora; and PERMA (Liu et al., 2026) combining two noise levels with two task scopes. We report accuracy for PersonaMem, PrefEval, and PERMA, and the F1 score for Per- sonaBench. See Appendix A for details. Baselines. We compare our framework against a diverse selection of baselines. LongContext feeds raw histories directly into the context win- dow, whereas RAG retrieves the top-krelevant seg- ments from a vector store. We also include three retrieval-based methods that dynamically main- tain an external memory bank: Mem0 (Chhikara et al., 2025), A-Mem (Xu et al., 2026b), and Light- Mem (Fang et al., 2026). Furthermore, we compare against three RL-based memory agents that learn memory evolution actions: MemAgent (Yu et al., 2026a), MEM-α(Wang et al., 2026), and Mem- Skill (Zhang et al., 2026a). For a fair comparison, all baselines use the same configurations. Implementation. We employ Qwen2.5-7B- Instruct (Qwen et al., 2025) as the backbone language model for most methods, whereas MEM-αutilizes Qwen3-4B and MemSkill uses Qwen3-Next-80B-A3B-Instruct.For retrieval, all-MiniLM-L6-v2 (Wang et al., 2020) extracts the top ten candidates. Our training set consists of 423 instances sampled from 70% of the PersonaMem 32k subset. To reduce computational overhead, training uses retrieved dialogues as context, whereas inference utilizes the full history, processing context in 4K-token chunks. Baselines are implemented using their public repositories; for a fair comparison, MemAgent, MEM-αand MemSkill are initialized from public checkpoints and fine-tuned on our 423 samples. All experi- ments run on four NVIDIA H800 GPUs, with shared hyperparameters detailed in Appendix C. 5.2 Main Results Overall Comparison with Baselines.As shown in Tab. 1, HiPS consistently demonstrates superior performance across all twelve evaluation settings. This performance gap indicates that learning an ex- plicit, dynamically evolving memory-management strategy is significantly more effective than relying on fixed context inclusion or manually designed retrieval heuristics. Specifically, the performance of the Long Context baseline degrades severely un- der noisy, long-horizon histories, falling to 20.74 on PersonaMem 128K. In contrast, HiPS main- tains stable personalized performance by utiliz- ing its distilled strategies to filter out irrelevant information during memory updates. Compared to explicit memory-bank baselines such as Mem0, A-Mem, and LightMem, HiPS delivers substan- tial and consistent improvements across both in- domain and out-of-domain tasks. Although rein- forcement learning-based memory agents, specif- ically MemAgent, MEM-α, and MemSkill, are competitive, they still lag behind the holistic per- formance of our framework. This consistent ad- vantage highlights the efficacy of our hierarchical co-evolution design, where universal strategy dis- PersonaMemPERMA Configuration 32K128KC-SC-MN-SN-M HiPS (Full)73.4962.0166.9556.5663.8356.30 w/o Flow71.0860.4945.3940.8743.1236.50 w/o Gate 69.2859.6351.6344.9952.6246.79 w/o USD66.2753.7061.2851.9360.1449.10 w/o PG67.4752.2260.7051.4159.4353.98 w/o PDD69.8856.0560.9953.9860.7151.41 Table 2: Ablation results on PersonaMem and PERMA. Cell color intensity indicates drop from Full (<3 to ≥15 p). Rows ordered by average drop. tillation induces a transferable global guideline, and persona delta distillation dynamically adapts to user-specific behavioral patterns. Generalization Across Settings.Across the per- formance detailed in Tab. 1, HiPS exhibits robust generalizability. It consistently outperforms base- lines on both in-domain tasks and out-of-domain benchmarks. This includes the explicit and implicit preference queries of PrefEval, the increasingly noisy contexts of PersonaBench, and the diverse domain crossings of PERMA. This strong gener- alization capability is driven by our hierarchical strategy decomposition, where universal strategy distillation establishes stable, globally shared mem- ory organizations, and persona delta distillation handles targeted adaptations for divergent users. Detailed performance comparisons across different user categories are provided in Appendix B. 5.3 Ablation Study Tab. 2 ablation results reveal a striking in-domain versus out-of-domain importance flip between Per- sonaMem and PERMA. In-domain, removing USD or PG causes the largest declines, reducing Person- aMem 128K accuracy from 62.01 to 53.70 and 52.22, respectively, confirming that universal rule quality is critical when distributions align. Un- der this setting, PDD yields a moderate drop to 56.05, while Gate and Flow reduce accuracy only to 59.63 and 60.49, as training personas are well- served byS u alone. Conversely, this ranking re- verses out-of-domain. Ablating Flow triggers the most severe degradation, reducing C-S accuracy from 66.95 to 45.39, an average PERMA decrease of 19.4 points, while removing Gate drops C-S accuracy to 51.63, averaging an 11.9-point loss. These drops show that cross-level rule migration and divergence-gated personalization are essential when universal rules do not transfer, especially 0100200300400 Training Step 0 1 2 3 4 5 6 7 8 9 S u Rule Count Established Tentative Supported Accuracy 50 55 60 65 70 75 Accuracy (%) 52.4% 56.7% 59.9% 67.4% 72.1% Figure 3: Evolution ofS u evidence distribution and accuracy throughout training on PersonaMem. in multi-domain settings such as C-M and N-M, where cross-domain interests stress-test persona diversity. In contrast, USD and PG matter less out- of-domain, dropping performance by only 5.3 and 4.5 points on average. Finally, ablating PDD causes consistent, moderate drops of 6.0 on PersonaMem 128K and 4.1 on average across PERMA. 5.4 Strategy Quality Analysis Evolutionary Dynamics of Universal Strategies. Fig. 3 tracesS u evidence distribution and accu- racy throughout training on PersonaMem 32K val- idation set. Starting from 5 seed rules (all sup- ported), the system hypothesizes new rules (appear- ing as tentative) and validates successful ones. The stacked area shows evidence of evolution: rules progressively mature from tentative to supported to established, with accuracy rising from 52.4% to 72.1% by convergence. The rule count oscil- lates before stabilizing at 6 established rules at step 480+, reflecting the create-validate-prune lifecycle. Qualitative Analysis of Personalized Deltas. Fig. 4 illustrates how the impact of∆ p varies across user profiles. For P15, the finance enthu- siast, a specialized financial logging rule yields an 18.6% point improvement by preventing granular data compression. Similarly, P2, the programmer, gains 17.9% points by preserving conflicting port- folio entries instead of overwriting them, while P0, the musician, gains 15.8% points through custom milestone tracking. Conversely, P6, the market- ing specialist, experiences a 3.3% point drop. Be- cause P6’s professional activities are already well- managed byS u , additional rules over-constrain model generation, validating our divergence-gating design to selectively apply personalization. P15 Finance Enthusiast Profile : Focuses on financial literacy, investment strategies, and budgeting When summarizing activities, including both frequency and impact details for financial literacy and investment strategies. P2 Programmer Profile : Python developer and investor. Manages Financial portfolios alongside software projects. When new financial information contradicts existing entries, preserve both versions but highlight discrepancies for review. P0 Musician Profile : Soft engineer and musician. Blends Pacific Islander music with digital production. When recording musical projects, including specific milestones and achievements. P6 Marketing Specialist Profile : Marketing specialist. Enjoys book clubs and professional development Preserve specific outcomes and metrics for professional activities. Figure 4: Per-persona impact of persona-specific strategies (∆ p ). We compare Full HiPS (with∆ p ) against w/o PDD (without ∆ 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 RAG52.4151.5860.8363.74 A-Mem 55.4255.9162.7465.93 ▼ Optimized w/ Qwen2.5-7B-Instruct HiPS (S u only)57.8356.7464.2567.18 HiPS (S u +∆ p ) 61.2459.8367.4570.16 ▼ Optimized w/ GPT-4o-mini HiPS (S u only) 56.4757.3565.1868.42 HiPS (S u +∆ p ) 60.0761.3568.1271.48 Table 3: Cross-model transferability of strategies (with- out RL policy). Strategies are distilled with one LLM and injected into others’ prompts for inference. 5.5 Cross-model Transfer Tab. 3 presents the transferability of the distilled strategies across various backbone LLMs. Both HiPS variants consistently outperform baselines such as RAG and A-Mem across all models, demonstrating that our strategy captures model- agnostic memory management principles rather than overfitting to a specific LLM. Moreover, the S u +∆ p configuration surpasses theS u -only base- line, confirming that hierarchical personalization provides complementary benefits across backbones. Notably, optimization viaGPT-4o-minigeneral- izes well, achieving strong results on three back- bones, includingGPT-5at 71.48 andGemini 2.5 flash at 68.12. These results confirm that HiPS produces portable strategies, enabling deployment across diverse backbone models. 5.6 Scaling Analysis Fig. 5 analyzes how our framework scales with longer dialogue context on PersonaMem (128K). As the dialogue tokens grow from 4K to 128K, 4K8K16K32K64K128K Dialogue Tokens 1000 1200 1400 1600 1800 Memory Bank Size (tokens) Memory Bank Size Memory Evolving Time 200 400 600 800 1000 1200 1400 1600 Memory Evolving Time (s) 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). the memory bank size increases from 1,000 to around 1,900 tokens, and the curve exhibits a clear trend: it rises steadily in the short-context regime and gradually flattens as the dialogue lengthens, which demonstrates that HiPS effectively consoli- dates preference-relevant information while filter- ing out redundant content to keep memory growth bounded. Meanwhile, the memory evolving time scales approximately linearly with dialogue length, ranging from roughly 140 to 1,700 seconds across the full spectrum. This indicates that the compu- tational overhead remains predictable and well- controlled, confirming that HiPS can efficiently handle ultra-long dialogues without incurring pro- hibitive memory or time costs. 6 Conclusion In this paper, we introduce HiPS, a framework that decouples memory management strategies into a universal baseline and a user-specific adaptive delta, with the partition updated using online evidence. Our empirical results verify that HiPS secures con- sistent performance gains over retrieval and RL- trained counterparts, with its advantages becoming pronounced as the dialogue context scales. By es- tablishing a continuous co-evolution loop between multi-tiered strategies and active policies, this work paves the way for developing robust and truly per- sonalized long-term memory systems for agents. Limitations While HiPS demonstrates robust improvements across diverse memory benchmarks, we identify several gentle limitations that offer promising av- enues for future research. First, our evaluation is primarily focused on English-language conversa- tional benchmarks. While the core algorithmic principles of strategy distillation and policy co- evolution are model-agnostic, memory manage- ment strategies in other typological languages may exhibit different syntactic structures or require lo- calized prompt adjustments. Testing our frame- work in multilingual and cross-lingual settings re- mains an open area for future work. Second, our experiments evaluate interaction histories spanning up to 1M tokens, which represents a medium-to- long timeline of conversational sessions. Over extremely long-term lifecycles, such as years of continuous daily interaction, a user’s fundamental baseline personality traits may undergo gradual, paradigm-level shifts. Handling such slow-moving, long-term personal baseline drift would require ad- ditional meta-distillation mechanisms to update the universal seed rules themselves occasionally. Acknowledgments This research was partially supported by the Na- tional Natural Science Foundation of China (Grants No.62406303). References Prateek Chhikara, Dev Khant, Saket Aryan, Taranjeet Singh, and Deshraj Yadav. 2025. Mem0: Building production-ready ai agents with scalable long-term memory. Preprint, arXiv:2504.19413. Jizhan Fang, Xinle Deng, Haoming Xu, Ziyan Jiang, Yuqi Tang, Ziwen Xu, Shumin Deng, Yunzhi Yao, Mengru Wang, Shuofei Qiao, Huajun Chen, and Ningyu Zhang. 2026. Lightmem: Lightweight and efficient memory-augmented generation. Preprint, arXiv:2510.18866. Pengyu Gao, Jinming Zhao, Xinyue Chen, and Long Yilin. 2025. An efficient context-dependent memory framework for LLM-centric agents. In Proceedings of the 2025 Conference of the Nations of the Amer- icas Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 3: Industry Track), pages 1055–1069, Albuquerque, New Mexico. Association for Computational Linguis- tics. Chuanrui Hu, Xingze Gao, Zuyi Zhou, Dannong Xu, Yi Bai, Xintong Li, Hui Zhang, Tong Li, Chong Zhang, Lidong Bing, and Yafeng Deng. 2026a. Ev- ermemos: A self-organizing memory operating sys- tem for structured long-horizon reasoning. Preprint, arXiv:2601.02163. Yuyang Hu, Shichun Liu, Yanwei Yue, Guibin Zhang, Boyang Liu, Fangyi Zhu, Jiahang Lin, Honglin Guo, Shihan Dou, Zhiheng Xi, Senjie Jin, Jiejun Tan, Yan- bin Yin, Jiongnan Liu, Zeyu Zhang, Zhongxiang Sun, Yutao Zhu, Hao Sun, Boci Peng, and 28 others. 2026b. Memory in the age of ai agents. Preprint, arXiv:2512.13564. Bowen Jiang, Zhuoqun Hao, Young-Min Cho, Bryan Li, Yuan Yuan, Sihao Chen, Lyle Ungar, Camillo J. Taylor, and Dan Roth. 2025. Know me, respond to me: Benchmarking llms for dynamic user profil- ing and personalized responses at scale. Preprint, arXiv:2504.14225. Zhiyu Li, Shichao Song, Hanyu Wang, Simin Niu, Ding Chen, Jiawei Yang, Chenyang Xi, Huayi Lai, Ji- hao Zhao, Yezhaohui Wang, Junpeng Ren, Zehao Lin, Jiahao Huo, Tianyi Chen, Kai Chen, Kehang Li, Zhiqiang Yin, Qingchen Yu, Bo Tang, and 3 others. 2025. Memos: An operating system for memory- augmented generation (mag) in large language mod- els. Preprint, arXiv:2505.22101. Shuochen Liu, Junyi Zhu, Long Shu, Junda Lin, Yuhao Chen, Haotian Zhang, Chao Zhang, Derong Xu, Jia Li, Bo Tang, Zhiyu Li, Feiyu Xiong, Enhong Chen, and Tong Xu. 2026. Perma: Benchmarking personalized memory agents via event-driven pref- erence and realistic task environments. Preprint, arXiv:2603.23231. Siru Ouyang, Jun Yan, I-Hung Hsu, Yanfei Chen, Ke Jiang, Zifeng Wang, Rujun Han, Long Le, Samira Daruki, Xiangru Tang, Vishy Tirumalashetty, George Lee, Mahsan Rofouei, Hangfei Lin, Jiawei Han, Chen-Yu Lee, and Tomas Pfister. 2026. Reasoning- bank: Scaling agent self-evolving with reasoning memory. In The Fourteenth International Confer- ence on Learning Representations. Charles Packer, Sarah Wooders, Kevin Lin, Vivian Fang, Shishir G. Patil, Ion Stoica, and Joseph E. Gonzalez. 2024. Memgpt: Towards llms as operating systems. Preprint, arXiv:2310.08560. Sriyash Poddar, Yanming Wan, Hamish Ivison, Ab- hishek Gupta, and Natasha Jaques. 2024. Person- alizing reinforcement learning from human feedback with variational preference learning. Advances in Neural Information Processing Systems, 37:52516– 52544. Qwen, An Yang, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chengyuan Li, Dayiheng Liu, Fei Huang, Haoran Wei, Huan Lin, Jian Yang, Jianhong Tu, Jianwei Zhang, Jianxin Yang, Jiaxi Yang, Jingren Zhou, Junyang Lin, and 24 others. 2025. Qwen2.5 technical report. Preprint, arXiv:2412.15115. Zhihong Shao, Peiyi Wang, Qihao Zhu, Runxin Xu, Junxiao Song, Xiao Bi, Haowei Zhang, Mingchuan Zhang, Y. K. Li, Y. Wu, and Daya Guo. 2024. Deepseekmath: Pushing the limits of mathemati- cal reasoning in open language models. Preprint, arXiv:2402.03300. Chuyi Tan, Peiwen Yuan, Xinglin Wang, Yiwei Li, Shaoxiong Feng, Yueqi Zhang, Jiayi Shi, Ji Zhang, Boyuan Pan, Yao Hu, and Kan Li. 2025a. Diagnos- ing and mitigating system bias in self-rewarding rl. Preprint, arXiv:2510.08977. Juntao Tan, Liangwei Yang, Zuxin Liu, Zhiwei Liu, Rithesh R N, Tulika Manoj Awalgaonkar, Jianguo Zhang, Weiran Yao, Ming Zhu, Shirley Kokane, Silvio Savarese, Huan Wang, Caiming Xiong, and Shelby Heinecke. 2025b. PersonaBench: Evaluating AI models on understanding personal information through accessing (synthetic) private user data. In Findings of the Association for Computational Lin- guistics: ACL 2025, pages 878–893, Vienna, Austria. Association for Computational Linguistics. Wenhui Wang, Furu Wei, Li Dong, Hangbo Bao, Nan Yang, and Ming Zhou. 2020. Minilm: Deep self-attention distillation for task-agnostic com- pression of pre-trained transformers.Preprint, arXiv:2002.10957. Yu Wang, Ryuichi Takanobu, Zhiqi Liang, Yuzhen Mao, Yuanzhe Hu, Julian McAuley, and Xiaojian Wu. 2026. MEM-$α$: LEARNING MEM- ORY CONSTRUCTION VIA REINFORCEMENT LEARNING. Derong Xu, Shuochen Liu, Pengfei Luo, Pengyue Jia, Yingyi Zhang, Yi Wen, Yimin Deng, Wenlin Zhang, Enhong Chen, Xiangyu Zhao, and Tong Xu. 2026a. Learning how and what to memorize: Cognition- inspired two-stage optimization for evolving memory. Preprint, arXiv:2605.00702. Wujiang Xu, Zujie Liang, Kai Mei, Hang Gao, Juntao Tan, and Yongfeng Zhang. 2026b. A-mem: Agentic memory for LLM agents. In The Thirty-ninth An- nual Conference on Neural Information Processing Systems. Sikuan Yan, Xiufeng Yang, Zuchao Huang, Ercong Nie, Zifeng Ding, Zonggen Li, Xiaowen Ma, Jinhe Bi, Kristian Kersting, Jeff Z Pan, and 1 others. 2025. Memory-r1: Enhancing large language model agents to manage and utilize memories via reinforcement learning. arXiv preprint arXiv:2508.19828. Hongli Yu, Tinghong Chen, Jiangtao Feng, Jiangjie Chen, Weinan Dai, Qiying Yu, Ya-Qin Zhang, Wei- Ying Ma, Jingjing Liu, Mingxuan Wang, and Hao Zhou. 2026a. Memagent: Reshaping long-context LLM with multi-conv RL-based memory agent. In The Fourteenth International Conference on Learn- ing Representations. Yi Yu, Liuyi Yao, Yuexiang Xie, Qingquan Tan, Jiaqi Feng, Yaliang Li, and Libing Wu. 2026b. Agen- tic memory: Learning unified long-term and short- term memory management for large language model agents. Preprint, arXiv:2601.01885. Haozhen Zhang, Quanyu Long, Jianzhu Bao, Tao Feng, Weizhi Zhang, Haodong Yue, and Wenya Wang. 2026a. Memskill: Learning and evolving memory skills for self-evolving agents. Preprint, arXiv:2602.02474. Yingyi Zhang, Junyi Li, Wenlin Zhang, Pengyue Jia, Xianneng Li, Yichao Wang, Derong Xu, Yi Wen, Huifeng Guo, Yong Liu, and Xiangyu Zhao. 2026b. Evoking user memory: Personalizing LLM via recollection-familiarity adaptive retrieval. In The Fourteenth International Conference on Learning Representations. Siyan Zhao, Mingyi Hong, Yang Liu, Devamanyu Haz- arika, and Kaixiang Lin. 2025. Do LLMs recognize your preferences? evaluating personalized preference following in LLMs. In The Thirteenth International Conference on Learning Representations. Zijian Zhou, Ao Qu, Zhaoxuan Wu, Sunghwan Kim, Alok Prakash, Daniela Rus, Bryan Kian Hsiang Low, and Paul Pu Liang. 2026. MEM1: Learning to syner- gize memory and reasoning for efficient long-horizon agents. In The Fourteenth International Conference on Learning Representations. A Datasets A.1 PersonaMem Benchmark PersonaMem (Jiang et al., 2025) is a large-scale benchmark designed to evaluate long-term person- alization in conversational LLMs. It features inter- action histories for 20 simulated personas, with each persona defined by rich static attributes, such as demographics, alongside dynamic traits and pref- erences that evolve across 15 diverse real-world task domains, including food recommendation, travel planning, and therapy consultation. For ev- ery persona, multi-session conversations are con- structed where the user engages with a chatbot via 7 types of in-situ queries that probe distinct personalization capabilities, such as recalling user facts, tracking preference evolution, and provid- ing preference-aligned suggestions. Each session consists of 15 to 30 user-assistant turns. These StatisticPersonaMem Tokens per history~32k ~128k~1M # QA pairs58927272674 # Sessions per history102060 Avg. # utterances167.1 758.3 3607.9 Table 4: Statistics of the PersonaMem dataset at dif- ferent context lengths. Token counts denote the ap- proximate total context length per interaction history; utterance counts are averaged over histories. StatisticValue Explicit queries1,000 Implicit queries1,000 Maximum inserted conversations24 Maximum inserted turns326 Avg. turns / conversation13.58 Total tokens108,102 Avg. tokens / conversation4,504.25 Table 5: Dataset statistics for PrefEval multiple-choice classification. histories are instantiated at three context scales by concatenating 10, 20, or 60 sessions, yielding ap- proximate context lengths of 32k, 128k, and 1M tokens, respectively. At evaluation time, models must select appropriate responses to user queries conditioned on the interaction history, testing their ability to adapt to dynamic user profiles. Tab. 4 summarizes the main statistics of PersonaMem. A.2 PrefEval Benchmark PrefEval (Zhao et al., 2025) is a long-context, multi- session benchmark designed to evaluate whether LLMs can infer, retrieve, and act on user pref- erences in realistic conversational settings. The benchmark emphasizes four core aspects: prefer- ence inference, long-context retrieval, preference following, and personalization proactiveness. The dataset comprises 1,000 unique preference-query pairs and spans 20 everyday topics grouped into seven domains: Entertainment, comprising shows, music and books, sports, and games; Travel, span- ning activities, restaurants, hotels, and transporta- tion; Lifestyle, including diet, beauty, fitness, and health; Shopping, encompassing home, fashion, motors, and technology; Education, covering ed- ucational resources and learning styles; Profes- sional Ownership; and Professional Work Style. PrefEval supports two evaluation formats, namely UserQueriesCorpusConv.AIE-com. 14811084233 243907884 3426451121 4468571140 5448459214 6409479141 Sum2635274229213 Table 6: Statistics of the PersonaBench subset across six users. Corpus is the sum of Conv., AI, and E-com.. a free-form generation setting and a 4-way multiple- choice classification setting where exactly one op- tion is consistent with the stated preference. To challenge long-range personalization capabilities, the benchmark inserts unrelated multi-session di- alogue turns between the preference revelation and the final query. In our experiments, we em- ploy 1,000 explicit and 1,000 implicit instances un- der the multiple-choice classification setting, with 50 intervening turns inserted as distractor context. Tab. 5 reports the summary statistics of this subset. A.3 PersonaBench Benchmark PersonaBench (Tan et al., 2025b) evaluates person- alized retrieval and question answering grounded in user-specific contexts. For each user, the dataset provides a heterogeneous personal corpus consist- ing of conversations with friends, denoted as Conv., dialogues with AI assistants, denoted as AI, and e-commerce purchase histories, denoted as E-com.. The evaluation queries are typically short and un- derspecified, requiring models to resolve implicit intent by grounding responses in evidence dis- tributed across historical interactions and behav- iors. This setup tests a model’s ability to align with diverse, user-dependent semantics under real- istic contextual ambiguity. Tab. 6 summarizes the per-user query counts and corpus statistics for the six-user subset used in our experiments. A.4 PERMA Benchmark PERMA (Liu et al., 2026) is an event-driven, long- context benchmark for evaluating whether per- sonalized memory agents can maintain, update, and synthesize dynamic persona states in realis- tic conversational environments, with an empha- sis on three aspects: task completion, preference consistency, and informational confidence. The dataset comprises 10 representative user profiles containing 2,166 fine-grained preference details across 10 countries, and spans 20 distinct domains, MetricP1P2P3P4P5P6P7P8P9P10Total User Demographics Age35-4425-3465+55-6435-4418-2425-3435-4425-3425-34– GenderMMFMMMMMFM– EducationGrad.Univ.–Univ.–Grad.Univ.–Grad.Univ.– CountryCAMXFIUSAUUKCHILRUBE– Interaction Statistics (Base Dataset: Clean / Noisy) # Interests17161515171313171612151 # Queries63626059635150636049580 # Events81818579827975818184808 # Dialogs356/364 340/358 373/379 347/352 348/370 349/357 321/329 350/364 359/362 365/375 3.5k/3.6k # Tokens (k)34/3432/3333/3631/3133/3332/3231/3132/3634/3933/34324/331 Style-aligned Long-Context Dataset # Events156141156139137107176941551271,388 # Dialogs1009848108487688880210157989698699,158 # Tokens (k)139.061.4174.088.9197.269.0151.162.7143.778.41,165.4 Table 7: Comprehensive statistics of the dataset across multiple dimensions. The table summarizes user demograph- ics 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. including Travel, Finance, Shopping, Entertain- ment, Messaging, and Calendar. PERMA sup- ports two evaluation formats: an 8-option multiple- choice question (MCQ) probing setting in which options are systematically ablated, and a multi- turn interactive setting driven by an LLM-based user simulator that terminates upon successful task completion. To stress realistic, long-horizon personalization, the benchmark structures interac- tions temporally through event-driven timelines and injects controlled within-session noise to sim- ulate real-world user erraticism. In our experi- ments, we evaluate memory agents across four key scenarios–Clean-Single, Clean-Multi, Noise- Single, and Noise-Multi–which span single- and multi-domain tasks under both standard dialogue histories and those perturbed with text-variability noise; comprehensive statistics of the dataset are reported in Tab. 7. B Comparison in Different Categories Comparison in Different Categories of Person- aMem. Tab. 8 reports category-wise results on PersonaMem under 32K and 128K interaction his- tories. Across both scales,HiPSachieves the best overall performance (73.49 at 32K; 62.01 at 128K), and the gains are concentrated on memory- dependent personalization abilities. On 32K his- tories,HiPSleads in Suggest ideas (51.85), Prefs evolve (84.44), Update reasons (89.66), Aligned recs (73.33), and New Scenarios (53.33), which jointly drives a clear margin over the strongest baselines (e.g., 73.49 vs. 64.45 for MemSkill). When scaling to 128K, Long Context degrades sharply overall (20.74), while memory-based meth- ods remain substantially stronger; within them, HiPSstays best-performing and ranks first on Re- call facts (68.42), Suggest ideas (44.02), Latest prefs (71.82), Prefs evolve (71.55), Update reasons (72.49), Aligned recs (54.73), and New Scenarios (44.13). In contrast, Recall facts is not a compara- tive strength forHiPSunder the 32K scale (71.43 vs. 72.87 for MemAgent), indicating that while simpler fact-retrieval mechanisms can be effective in shorter horizons,HiPS’s architectural advantage lies in tracking and applying evolving, structured preferences as the context scales up. Overall, the category-wise gains suggest that explicitly structur- ing memory operations and then selecting what to keep is most effective for preference-heavy queries, and the advantage becomes more pronounced as the interaction history grows longer. Comparison in Different Categories of PrefE- val. Tab. 9 breaks down PrefEval performance by domain under Explicit and Implicit preference settings. Under Explicit Memory,HiPSattains the best overall accuracy (89.20) and shows consis- tently strong gains on preference-heavy domains, ranking first across all evaluated topics, includ- ing Travel (90.20), Entertain (94.20), Lifestyle (91.70), Shop (84.10), Education (86.30), Profes- sional (85.20), and Pet (78.10). This comprehen- sive lead demonstrates thatHiPScan robustly ex- tract and apply directly stated constraints. Under Implicit Memory, the task becomes more challeng- ing for all methods, yetHiPSagain leads overall (69.40) and improves most clearly on domains that require inferring latent preferences from context, such as Travel (71.90), Lifestyle (71.60), Shop (64.50), and Professional (63.40). A notable ex- ception is Education, whereHiPS(65.10) slightly trails MemSkill (65.80), suggesting that certain implicit reasoning contexts may occasionally ben- efit from alternative skill-driven memory-update pipelines. Compared with memory-bank baselines (Mem0/A-Mem/LightMem), the advantage ofHiPS is broad across domains in both settings, indicating that it better resists long-range distractors and pre- serves preference-relevant signals. Overall, these domain-wise results reflect PrefEval’s construction: the inserted unrelated turns make long-range pref- erence retrieval and faithful preference following the main bottlenecks, andHiPSimproves most on domains where precise preference identification and consistent application are essential. C Hyperparameters Tab. 10 lists all HiPS hyperparameters and their default values. Hyperparameter Sensitivity Analysis To eval- uate how key hyperparameters in HiPS influence personalization accuracy, we conduct a parame- ter sensitivity analysis on the PersonaMem 128K benchmark. As shown in the sensitivity curves, the trends demonstrate clear trade-offs across all key configurations: The distillation frequencies, K 1 andK 2 , govern the update speed of universal and personalized rules. For the USD frequencyK 1 , the optimal performance is achieved at 20 steps. Too frequent distillation, such as every 5 steps, in- troduces short-term interaction noise, while too infrequent updates, such as every 80 steps, fail to adapt in a timely manner. For the PDD frequency K 2 , the peak occurs at 10 steps, indicating that per- sonalized strategies require more frequent updates than universal rules to capture rapid user-specific behavioral shifts. Regarding the strategy token budgetB, performance improves sharply as the budget increases from 80 tokens, peaking around 200 to 300 tokens. Although 300 tokens yields a marginally higher accuracy, we select 200 as the de- fault value to balance strategy expressiveness with prompt token efficiency. The divergence threshold θ div controls personalization gating. Low values like 0.1 trigger excessive, unnecessary personaliza- tion that introduces content-level noise, whereas high values like 0.5 over-filter users, preventing divergent individuals from receiving tailored rules. The optimal threshold is 0.3. The flow threshold θ flow regulates rule migration and exhibits robust performance across a wide range of values, reach- ing its peak at 0.6. The validation and revision thresholds,θ val andθ rev , peak at 0.1 and 0.02, re- spectively. These values indicate that a moderately conservative update policy is optimal to prevent the premature promotion of spurious rules while ensur- ing the timely pruning of contradicted guidelines. Finally, the quality weightλcontrols the balance between task reward and strategy adherence, peak- ing at 0.3. Removing the adherence signal entirely, whereλis 0.0, reduces performance, while over- weighting adherence at 0.8 degrades task success by forcing compliance over objective outcomes. D Seed Rules forS u The five seed rules used to initializeS u . All seeds start at the Supported evidence level and must earn Established status through training; they can be revised or pruned if contradicted by evidence. 1. Structure: When processing chunks, organize memory into labeled sections such as Identity, Preferences, and Activities. 2.Conflict resolution: When new information contradicts existing memory, mark the old en- try as deprecated with a reason and keep both. 3.Relevance filtering: When encountering one- off details like transient locations or momen- tary moods, omit them unless explicitly stated as long-term. 4. Compression: When duplicate or redundant entries appear across sections, merge them into a single consolidated entry. 5.Evidence-bound: When updating memory, only store information directly supported by the current conversation chunk. E Evidence Level Lifecycle Each rule carries an evidence level that reflects its validation history across distillation cycles: Method Recall facts Suggest ideas Latest prefs Prefs evolve Update reasons Aligned recs New Scenarios Overall ▼ 32K memory corpus data Long Context41.3822.22-42.2272.4133.3326.6742.17 RAG37.9322.22-64.4486.2160.0033.3352.41 Mem047.9319.41-46.6179.5857.0442.7548.53 A-Mem 45.7118.52-73.3379.3166.6733.3355.42 LightMem47.0612.90-66.1975.7638.1824.5652.41 MemAgent72.8723.66-69.7883.8450.9138.6059.93 Mem-α45.7129.63-71.1186.2166.6740.0058.43 MemSkill57.1440.74-73.3389.6666.6746.6764.45 HiPS 71.4351.85-84.4489.6673.3353.3373.49 ▼ 128K memory corpus data Long Context9.2622.1515.0828.8736.4922.0216.6720.74 RAG 53.7020.2533.7352.5859.4639.4536.3638.02 Mem056.2520.4937.7755.0957.2041.6429.1339.67 A-Mem64.8118.9935.7154.6462.1640.3737.8839.88 LightMem 30.9914.6735.3361.2964.3135.8228.1736.74 MemAgent56.1427.8061.6664.2266.1740.6934.7450.86 Mem-α 59.2630.3850.4060.8264.8645.8739.3948.18 MemSkill64.8137.3461.5167.0168.9250.4642.4255.37 HiPS68.4244.0271.8271.5572.4954.7344.1362.01 Table 8: Category-wise accuracy (%) on PersonaMem under 32K and 128K interaction histories. “–” indicates the category is not available in the dataset. MethodTravelEntertainLifestyleShopEducationProfessionalPetOverall ▼ Explicit Memory Long Context33.9042.3037.8024.1024.7023.708.0032.40 RAG 48.5055.9052.4040.3040.6038.0025.9047.40 Mem0 60.1064.0062.3049.7053.2050.2038.7057.60 A-Mem 63.9069.2065.6056.7056.8055.5043.3062.30 LightMem66.1070.9067.2059.3057.5056.3047.0064.20 MemAgent78.6081.1082.4071.6073.2073.1065.2077.40 Mem-α78.1080.2079.4069.8071.6070.7060.6075.70 MemSkill84.1086.5085.0079.1079.3078.5068.8082.60 HiPS 90.2094.2091.7084.1086.3085.2078.1089.20 ▼ Implicit Memory Long Context28.0032.5031.1019.3020.3015.608.0025.90 RAG 33.5042.4036.2024.2027.6024.109.4032.40 Mem048.3053.2051.5039.2040.3039.5026.4046.40 A-Mem54.8058.7056.6046.6049.2045.0034.8052.80 LightMem57.5060.4059.6048.6047.5047.5037.3054.80 MemAgent67.1069.0067.5058.4060.7059.0050.3064.40 Mem-α 64.6071.1065.3056.6059.6055.4045.6063.10 MemSkill 68.4073.0071.5063.6065.8062.7054.3068.40 HiPS71.9074.2071.6064.5065.1063.4055.3069.40 Table 9: Domain-wise accuracy (%) on PrefEval multiple-choice classification under Explicit vs. Implicit preference. •Tentative: Newly hypothesized, awaiting con- firmation. Not injected into the prompt. • Supported: Validated by at least one distilla- tion cycle. Eligible for prompt injection. •Established: Consistently validated across multiple cycles. Highest priority for injection. Rules are promoted when trajectory data shows they predict task success (predictive gain exceeds θ val ), and demoted when contradicted by evidence (predictive gain falls belowθ rev ). Stale Tentative rules that receive no validation withinT stale consec- utive cycles are pruned. F USD and PDD Prompts USD prompt. The USD prompt instructs the LLM to compare high-reward and low-reward tra- jectories across users and identify management pat- 510204080 USD Frequency (steps) 58 59 60 61 62 63 Accuracy (%) K 1 510204080 PDD Frequency (steps) Accuracy (%) K 2 80120200300400 Max Active Tokens Accuracy (%) B 0.10.20.30.40.5 Divergence Threshold Accuracy (%) div 0.40.60.81.0 Flow Threshold 58 59 60 61 62 63 flow 0.020.050.10.20.3 IG Validate Threshold val 0.0050.010.020.050.1 IG Revise Threshold rev 0.00.10.30.50.8 Quality Weight 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 frequencyK 1 , PDD frequencyK 2 , strategy token budgetB, divergence thresholdθ div , flow threshold θ flow , validation threshold θ val , revision threshold θ rev , and quality weight λ. HyperparameterSymbolValue Universal Strategy Distillation (USD) USD distillation frequencyK 1 20 Distillation trajectories per cyclek5 Trajectory buffer size—50 Promotion thresholdθ val 0.1 Demotion thresholdθ rev 0.02 Stale pruning cyclesT stale 3 Persona Delta Distillation (PDD) PDD distillation frequencyK 2 10 Gating thresholdθ div 0.3 Rule Budget & Flow Total token budgetB200 Minimum budget floor—0.2B Migration thresholdθ flow 0.6 Policy Optimization Reward weight (adherence)λ0.3 Table 10: Hyperparameter settings used for HiPS strat- egy co-evolution. terns that generalize. To focus the LLM on manage- ment patterns rather than content, the prompt ex- plicitly instructs: “IGNORE user-specific content (names, hobbies, topics). Focus only on how the memory was ORGANIZED and MANAGED.” The prompt is enriched with behavioral feature statistics (verbosity, structural organization, change tracking) and their contrastive attribution across high- and low-reward groups. Key output constraints: (1) structured diffs only (validate/hypothesize/revise), (2) rules follow “[Label]: When [condition], [ac- tion]” format, (3) rules describe management ac- tions, not topic-level content preferences. When a hypothesized rule overlaps an existing entry (word overlap> 0.7), it is treated as a validation rather than a new hypothesis, preventing rule explosion. PDD prompt. The PDD prompt presentsS u as anchoring context and instructs the LLM to iden- tify management behaviors for a specific user that differ from the universal rules. Key constraints be- yond USD: (1) rules must describe management actions, not content topics, (2) rule conditions must reference observable behavioral patterns, not user identities or domain names. Concretely, a rule like “preserve full context for financial discussions rather than summarizing” is valid because the ac- tion is about how to handle information; “the user is interested in finance” is not, because it states what the user cares about without prescribing a manage- ment decision. During training, rules are persona- specific; behavior-conditioning enables generaliza- tion to unseen users at inference time. G Compliance Estimation Both the predictive gain (PG, Eq. 4) and the ad- herence reward (R follow , Eq. 7) require estimating whether a trajectory complies with a given rule. We use a lightweight keyword-based heuristic: for each rule, we define a set of trigger keywords de- rived from the rule’s condition and action clauses; a trajectory segment is marked as compliant (r + ) if it contains at least one keyword from the action set and the corresponding condition keywords appear in the preceding context, and as non-compliant (r − ) otherwise. This heuristic is intentionally simple to avoid adding LLM calls during training. While im- perfect, PG is used for relative ranking (submodular Universal Strategy Distillation Template You are analyzing a memory management agent that maintains a structured user memory profile across multiple conversation sessions. At each episode, the agent reads conversation chunks one by one, updates a running user memory, and then answers a question about the user based solely on this memory. ## Current universal management rules: current_rules ## New evidence from recent episodes across DIFFERENT users: ### High-reward traces, where the memory led to correct answers: top_traces ### Low-reward traces, where the memory led to wrong answers: bottom_traces ## Instructions Compare the traces. IGNORE user-specific content, such as names, hobbies, or topics. Focus ONLY on how the memory was ORGANIZED and MANAGED. Good rules describe management ACTIONS, for example: - "Use labeled sections for different information types" - "Mark changed preferences as deprecated instead of deleting" Bad rules describe topics, for example: - "Prioritize cooking preferences", which represents content rather than management Output ONLY a JSON object with no other text: - "validate": list of rule indices as integers supported by evidence - "hypothesize": list of 1 to 2 NEW management rules, requiring fewer than 20 words per rule - "revise": list of rule indices as integers that conflict with evidence Example: "validate": [0, 2], "hypothesize": ["Use deprecated markers when preferences change"], "revise": [1] If no changes needed: "validate": [], "hypothesize": [], "revise": [] Figure 7: The complete prompt template used for Universal Strategy Distillation. selection, budget allocation) and threshold-based gating (divergence, auto-promotion), both of which are robust to systematic bias in the estimate: a uni- form upward shift in all PG values preserves their relative ordering and does not change which users exceed the divergence threshold. The adherence rewardR follow serves as a dense auxiliary signal alongside the sparse task rewardR ans ; its weight λ = 0.3is small, limiting the influence of any single compliance misjudgment. H Algorithm Algorithm 1 outlines the complete co-evolutionary training procedure for our framework. At each step, the system dynamically constructs an active strat- egy by merging the universal baselineS u and the personalized delta∆ p through budgeted submodu- lar selection. This active strategy guides the policy rollouts to generate interaction trajectories. After computing both the task correctness and strategy adherence rewards, the system strictly separates their downstream application. It updates the trajec- tory distillation buffer using exclusively the task reward to prevent self-validation loops, while uti- Persona Delta Distillation Template You are analyzing a memory management agent serving ONE specific user. The agent already follows these universal management rules: universal_rules Current personalized rules for this user, where indices refer to this list: current_delta ## New evidence from this user's recent episodes: ### High-reward traces, where the memory worked well: top_traces ### Low-reward traces, where the memory was less useful: bottom_traces ## Instructions Identify management behaviors specific to THIS user that DIFFER from the universal rules. Do NOT repeat any universal rule. Each rule must describe a MANAGEMENT ACTION, not a topic: WRONG: "Prioritize financial interests" RIGHT: "Preserve full preference change history instead of keeping only the latest" Output ONLY a JSON object with no other text: - "validate": list of rule indices as integers from personalized rules that are confirmed - "hypothesize": list of 1 to 2 NEW user-specific management rules, requiring fewer than 20 words per rule - "revise": list of rule indices as integers that conflict with evidence Example: "validate": [0], "hypothesize": ["Keep full preference change history with reasons"], "revise": [] If no changes needed: "validate": [], "hypothesize": [], "revise": [] Figure 8: The complete prompt template used for Persona Delta Distillation. lizing the combined reward to execute the GRPO policy update. To ensure continuous strategy refine- ment, the framework periodically triggers its dual distillation modules. The system executes Persona Delta Distillation everyK 2 steps to generate adap- tive rules for highly divergent users, and it invokes Universal Strategy Distillation everyK 1 steps to abstract shared principles and elevate widely suc- cessful personalized rules into the global standard. Algorithm 1 HiPS: Hierarchical Strategy-Policy Co-Evolution Require:Policyπ θ , seed rulesS u , empty∆ p , bufferB 1: for each training step t do 2:SelectS p = S u ∪ ∆ p via submodular opti- mization (Eq. 6) 3:Inject S p into rollout prompts 4:Collect trajectoriesτ; computeR ans , R follow 5:Update bufferB with (m K ,R ans ,t) // task reward only 6: GRPO update onπ θ usingR ans +λ·R follow 7:if t mod K 2 = 0 then 8:Recompute PG for all rules; auto-promote / demote by thresholds 9: PDD: evolve∆ p for users withD(p) ≥ θ div (Eq. 5) 10:end if 11:if t mod K 1 = 0 then 12: USD: evolveS u from cross-persona tra- jectories 13:Generalize frequent∆ p rules→ S u ; con- solidate duplicates 14:end if 15: end for