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.
Figures & tables
Figure 1. Comparison between (top) existing user-agnostic storage and (bottom) our proposed method driven by user persona. Comparison between (top) existing user-agnostic storage and (bottom) our proposed method driven by user persona.
Figure 2. The system architecture and operational workflow of HMO.
Method
Retriever
Rec@5 ↑
NDCG@5 ↑
Acc. ↑
No History ( Tan et al., 2025 )
-
-
-
0.0
Long Context ( Tan et al., 2025 )
-
-
-
57.4
RAG ( Tan et al., 2025 )
Stella
59.2
-
61.4
RAG ( Tan et al., 2025 )
GTE
62.4
-
63.6
MemoryBank ( Tan et al., 2025 )
Specific1
58.6
-
59.6
LD-Agent ( Tan et al., 2025 )
Specific2
56.8
-
59.2
Table 1. Comparison of long-term memory performance on the LongMemEval dataset. Our proposed HMO outperforms existing RAG-based and memory-augmented agents in both retrieval recall and reasoning accuracy.
Figure 3. Demonstration of HMO application on OpenClaw, with a simple and informative user interface.
Method
Overall F1 ↑
A-Mem ( Liu et al., 2025 )
23.30
MemoryBank ( Liu et al., 2025 )
31.42
Mem0 ( Liu et al., 2025 )
42.13
MemoryOS ( Liu et al., 2025 )
42.84
MemVerse ( Liu et al., 2025 )
43.44
HMO (Ours)
45.65
Table 2. Overall F1 performance on the LoCoMo benchmark using the GPT-4o-mini model.
Configuration
Rec@5 ↑
NDCG@5 ↑
Acc. ↑
Time (m) ↓
Graph-based
75.9
86.6
88.0
448.30
w/o Tier 1
80.4
85.8
83.1
21.53
w/o Tier 1, 2
85.9
89.7
88.8
56.05
w/ Tier 1, 2
81.1
85.6
86.4
12.88
Table 3. Ablation results on LongMemEval.
Strategy
Recall ↑
Acc. ↑
Build (s) ↓
Inf. (min) ↓
MLLM-Scoring-only
80.7
86.1
1.63
15.05
Similarity-only
83.3
86.7
0.15
87.58
HMO (Hybrid)
81.1
86.4
0.19
12.88
Table 4. Efficiency comparison of memory update strategies.
Appendix figures & tables5 assets
Supplementary material from the paper’s appendix.
Appendix
Method
Answer Prediction (F1)
Single-Hop
Multi-Hop
Temporal
Open-Domain
Overall
A-Mem ( Xu et al., 2025a )
44.65
27.02
45.85
12.14
23.30
MemoryBank ( Xu et al., 2025a )
41.04
26.65
25.52
9.15
31.42
Mem0 ( Chhikara et al., 2025 )
38.72
28.64
48.93
47.65
42.13
MemoryOS ( Kang et al., 2025b )
35.27
41.15
20.02
48.62
42.84
MemVerse ( Liu et al., 2025 )
46.11
29.69
30.84
49.23
43.44
Appendix
Table 5. Detailed Performance Breakdown on the LoCoMo Benchmark (GPT-4o-mini). Performance is measured by F1 scores across Single-Hop, Multi-Hop, Temporal, and Open-Domain categories. The best results are highlighted in blue bold, and the second-best results are marked in cyan.
Initial Importance Scoring Evaluator E
Role
You are an expert Cognitive Memory Engine. Your task is to evaluate a new interaction record between a User and an AI Agent to determine its “Initial Importance” ( Im ) for long-term retention.
Context
∙ User Persona ( P ): {{user_persona}}
∙ Interaction Record ( m ): {{memory_content}}
Scoring Dimensions
Appendix
Table 6. Prompt template for MLLM-based initial importance scoring.
LLM-as-a-judge Evaluation Prompt
Role
You are an expert language model evaluator. I will provide you with a question, a ground-truth answer, and a model-generated response. Your task is to determine whether the response correctly answers the question by following these evaluation rules:
Evaluation Rules
∙ Answer Yes if the response contains or directly matches the correct answer.
∙ Answer Yes if the response includes all necessary intermediate steps leading to the correct answer.
∙ Answer No if the response provides only a partial answer or omits essential information.
Appendix
Table 7. Prompt template for LLM-as-a-judge evaluation.
Figure 4. System interface demonstration (example 1).
Figure 5. System interface demonstration (example 2).
Large language models (LLMs) demonstrate strong reasoning and generation abilities, but their fixed context windows limit long-term information accumulation and reuse across multi-session interactions. Existing memory-augmented systems often construct memory in a coarse and unstable manner, relying on inefficient memory representations or unstable unconstrained updates. To address these challenges, we propose AtomMem, a long-term memory system designed for value-dense storage and stable memory evolution. AtomMem introduces a Fact Executor, which selectively extracts high value atomic facts from long form interactions to serve as highly efficient memory representations. Subsequently, AtomMem organizes these facts into hierarchical event structures and temporal profiles, capturing coherent episodic contexts and tracking dynamically evolving user attributes over time. During retrieval, the system activates an associative memory graph to connect fragmented memories. Experiments on the LoCoMo benchmark confirm that AtomMem achieves state-of-the-art performance across various reasoning tasks, offering a scalable and economically viable solution for deploying intelligent personalized agents.
Yanyu Yao, Shangze Li, Zhi Zheng +4
State Key Laboratory of Cognitive Intelligence, University of Science and Technology of China, Hefei, China · Anhui University, Hefei, China
Long-term memory is crucial for personalized responses and long-horizon agent interactions. Existing methods often rely on LLMs to compress or rewrite dialogue histories and use the transformed memories as retrieval evidence. Despite the progress in organizing fragmented contexts, two major drawbacks persist: (1) information loss from compression, which discards fine-grained but later useful details, and (2) semantic drift from rewriting, which erodes the original tone and situated context. In this work, we propose a novel Human-profile Enhanced Retrieval Optimization framework for long-term agent memory (HERO). Specifically, HERO converts the dialogue history into a traceable heterogeneous memory graph that preserves raw dialogue text as evidence for reasoning, thereby mitigating information loss. For retrieval, HERO extracts initial anchors from the current query and incorporates human profiles via an iterative graph traversal; these anchors and profiles provide guidance signals that adaptively activate the most informative regions of the graph. Experiments on two benchmark datasets show that HERO outperforms strong baselines on both factual and personalized reasoning, while providing more faithful access to raw dialogue evidence.
Yuanhua Lin, Yile Li, Zhiyuan Zhao +2
China Mobile Information Technology Co., Ltd., Beijing, China
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.
Yupeng Han, Shuochen Liu, Kai Zhang +3
State Key Laboratory of Cognitive Intelligence, University of Science and Technology of China