Source-linked AI summary
HERO: Human-profile Enhanced Retrieval Optimization Framework for Long-term Agent Memory
Yuanhua Lin, Yile Li, Zhiyuan Zhao, Jing Shang, Jian Sun
TL;DR
Existing long-term memory methods risk losing useful details and weakening situated dialogue context through compression and rewriting. HERO preserves raw dialogue in a heterogeneous memory graph, uses query anchors and human profiles for iterative retrieval, and outperforms strong baselines on factual and personalized reasoning across two benchmarks.
Problem
Compression can discard later-useful dialogue details, while rewriting can weaken original wording, affect, speaker attribution, and situated context.
Method
HERO preserves raw dialogue as episodic evidence in a heterogeneous memory graph and uses query anchors, profiles, and iterative graph traversal to guide retrieval.
Results
HERO outperforms strong baselines on factual and personalized reasoning, achieving 56.06% F1 on LoCoMo and 70.63% accuracy on PERSONAMEM.
Takeaways & Limitations
Preserving raw episodic context with structured retrieval supports factually faithful reasoning, personalized responses, and cross-session reasoning.
Takeaways & Limitations
HERO lacks explicit memory lifecycle mechanisms, and inaccurate LLM-derived profile cues may affect retrieval trajectories.
Abstract
from arXiv · showhide
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.
1 Introduction
Existing long-term memory methods can lose fine-grained dialogue details and introduce semantic drift, motivating HERO’s graph-based, profile-guided retrieval design. HERO preserves raw episodic evidence while using iterative cues to retrieve relevant facts and dependencies, outperforming strong baselines on two benchmarks.
- Compression can omit later-useful temporal references, causal conditions, numerical facts, and subtle preference changes.
- Rewriting dialogue into summaries or profiles may weaken original wording, affect, speaker attribution, and situated context.
- HERO constructs a heterogeneous memory graph that preserves raw dialogue traces as episodic evidence while profile nodes guide activation rather than replace memories.The graph connects Episodic Traces, Episodic Units, and Episodic Cues through structural, semantic, and temporal links.
- HERO uses iterative cue activation across query, episodic-unit, and profile levels to retrieve explicit memories and trace cross-episodic dependencies.Human profiles first guide cue reconstruction and query expansion before path-based graph retrieval.
- 56.06% F1 on LoCoMo and 70.63% accuracy on PERSONAMEM show HERO outperforming strong baselines on factual and personalized reasoning tasks.
2 Related Work
Long-term memory research includes summary-based and graph-based approaches. HERO addresses the information loss and semantic distortion associated with rewriting by preserving original text in a heterogeneous graph and using profiles and queries to guide retrieval.
- Summary-based methods refine original interactions into hierarchical summaries or structured facts to bypass context-window limitations.Representative systems use LLM-based compression or memory lifecycle management, but rewriting can cause information loss and semantic drift.
- Graph-based methods extract structured knowledge and perform path-based retrieval, supporting multi-hop reasoning and temporal handling of knowledge conflicts.
- HERO differs by constructing a heterogeneous memory graph faithful to original text and using iterative human-profile and query cues to guide personalized responses.
3 Methodology
HERO constructs a traceable heterogeneous memory graph from dialogue and profiles, then retrieves evidence through query anchors, iterative cue expansion, and graph-based relevance propagation.
- Problem formulation: HERO models dialogue sessions and a latent human profile, retrieving a memory subset that maximizes the generation probability of the target response.The backbone LLM parameters are denoted by ΦLLM.
- Memory Construction: HERO separates memory construction into an Episodic Layer for raw events and a Profile Layer for distilled user attributes.The framework organizes multi-granular conversational experiences into a coarse-to-fine heterogeneous graph.
- Memory Construction: The Episodic Layer contains traces, sentence-level units, and cues, preserving original text, timestamps, sequence indices, contextual constraints, and normalized entities.Episodic Traces represent local conversation scenarios, while Episodic Units carry detailed evidence and Episodic Cues act as associative bridges.
- Memory Construction: Profile Insights distill facts, insights, and traits from episodic traces, while Profile Cues link these abstractions back to the shared cue space.The global cue set is the union of episodic and profile cues.
- Graph Construction: The graph uses structural inclusion, weighted semantic indexing, and temporal chaining to connect cues, traces, units, profiles, and adjacent events.Semantic-index edge weights are based on occurrence frequency, while temporal edges preserve episodic order.
- Memory Retrieval: For retrieval, HERO handles factual and personalized questions using query entities, profile-related cues, iterative context expansion, and global relevance propagation.The three phases are Initial Anchor Identification, Iterative Context Expansion, and Global Relevance Propagation and Evidence Ranking.
- Memory Retrieval: Initial anchors match query entities to cue nodes, after which cue–context–cue traversal transfers activation across iterations before evidence ranking.The final ranking combines activation signals with dense retrieval relevance and uses masked Personalized PageRank before returning top-k episodic traces.
4 Experiments
HERO is evaluated on factual retrieval, cross-session reasoning, personalized reasoning, retrieval efficiency, component ablations, and hyperparameter sensitivity. It achieves strong task performance while preserving raw dialogue evidence and maintaining low retrieval overhead.
- Overall Performance: HERO achieves 87.99% accuracy, 56.06% F1, and 48.38% BLEU-1 on LoCoMo, leading all reported metrics.Its answers are grounded directly in raw dialogue text, improving both correctness and wording fidelity.
- Retrieval Quality: 0.938 overall Recall@10 shows that HERO locates relevant raw evidence in most LoCoMo cases.Open-domain questions have lower recall because required commonsense or world knowledge may not appear in the dialogue history.
- Cross-session Reasoning: 85.11% accuracy on LoCoMo Multi-hop tasks demonstrates HERO’s ability to link scattered clues across temporally distant sessions.Cue-based graph traversal connects episodic traces through entities and supports multi-hop retrieval of long-range dependencies.
- Personalized Reasoning: 70.63% accuracy on PERSONAMEM is the best overall result, with strong performance on preference-aligned recommendations and preference evolution.Reported category scores include 80.00% for recommendations, 87.88% for revisiting update reasons, and 72.66% for tracking full preference evolution.
- Efficiency: HERO attains its highest F1 with 248 average retrieval tokens and latency below one second.Semantic linking is moved to construction and profile maintenance, leaving lightweight cue extraction and graph traversal during retrieval.
- Ablation and Robustness: Ablations and sensitivity analyses show that profile guidance, cue expansion, and moderate threshold changes materially affect or preserve HERO’s behavior.The reported LoCoMo ranges remain 0.801–0.829 accuracy and 0.546–0.560 F1 under tested hyperparameter changes.
5 Conclusion
HERO addresses long-term personalized memory through an episodic-centered heterogeneous graph and iterative dual-cue activation. Across two benchmarks, it reports stronger personalized responses, accurate retrieval, and robust cross-session reasoning while preserving raw episodic context.
- Conclusion: HERO combines an episodic-centered heterogeneous memory graph with iterative dual-cue activation for personalized, factually faithful reasoning.The framework preserves raw episodic context while using structured retrieval to support long-term interactions.
- Conclusion: Experiments on two benchmark datasets demonstrate performance gains, accurate retrieval capabilities, and robust cross-session reasoning.The conclusion presents preserving raw episodic context with structured retrieval as a promising direction for long-term agent memory.
Limitations
HERO’s current implementation lacks explicit memory lifecycle controls and leaves several robustness issues for future work. These limitations concern trace maintenance, profile reliability, and long-running deployment.
- Long-running deployments may require additional mechanisms to manage the evolution and removal of raw episodic traces.The identified gap follows from HERO’s temporal chaining and semantic cue links without an explicit lifecycle mechanism.
- HERO does not include explicit time-aware decay, stale-trace pruning, or user-controlled deletion for stored episodic memories.Future work could add these strategies as lightweight local graph updates.
- Inaccurate LLM-extracted profile cues may affect the retrieval trajectory even though profiles are not directly used as generation evidence.The paper suggests confidence calibration and timestamp-aware profile updates as future improvements.
Ethical Considerations
The paper highlights risks from inaccurate stored or retrieved memories and recommends explicit data governance for real-world HERO deployments. These safeguards include deletion control and access transparency.
- Real-world HERO deployments should adopt explicit data governance policies, including user-controlled memory deletion and access transparency.The recommendation is framed as consistent with applicable data protection regulations.
- Inaccurate stored or retrieved memories could lead to misleading responses.HERO reduces this risk by grounding generation in raw, unmodified dialogue rather than rewritten summaries.
A.1 Experimental Settings
The experiments use fixed backbone and embedding models, standardized prompts and evaluation conditions, and benchmark-specific retrieval settings. Ablations use a smaller backbone to reduce computational cost.
- Unless otherwise specified, experiments use Qwen2.5-72B-Instruct for memory construction and answer generation and Qwen3-Embedding-8B for embeddings.All experimental LLMs are accessed through commercial API services.
- Ablation and hyperparameter analyses use Qwen3-30B-A3B to reduce computational cost.These results measure relative component contributions rather than absolute scores comparable to the main results.
- All methods use identical prompts and LLMs for answer generation and evaluation.EverMemOS and Mem0 follow the standardized EverMemOS evaluation pipeline for preprocessing and metric computation.
- Mem0’s main comparison uses its official hosted version as the strongest available implementation.
- Both benchmarks use chunk size 128, with retrieval top-k set to 10 on LoCoMo and 40 on PersonaMem.
A.2 Effect of Backbone LLMs
HERO remains competitive across backbone LLMs on LoCoMo, slightly exceeding EverMemOS under GPT-4o-mini and achieving strong accuracy with Qwen backbones. The reported results suggest its gains are not tied to one generation model.
- 0.870 accuracy under GPT-4o-mini slightly exceeds EverMemOS’s 0.868 on LoCoMo overall accuracy.The comparison is reported in Table 6.
- HERO’s backbone analysis indicates that its gains are not tied to a single generation model.The evidence comes from comparisons across GPT-4o-mini and two Qwen backbones.
- 0.880 accuracy with Qwen2.5-72B-Instruct and 0.866 with Qwen3-30B-A3B show stable HERO performance across Qwen backbones.
A.3 Selection of Retrieval Top-k
HERO remains stable across tested retrieval top-k values, while its graph representation has modest per-user memory and incremental construction costs.
- A.3 Selection of Retrieval Top-k: HERO remains stable across different top-k values on both LoCoMo and PERSONAMEM.On LoCoMo, top-k = 10 performs best; on PERSONAMEM, performance is stable from top-k = 30 to top-k = 60, with top-k = 40 used by default.
- A.4 Scalability and Construction Cost: The memory footprint stays below 3.4 MB per user across HERO’s evaluated graph configurations.The graph statistics distinguish trace, cue, and profile nodes.
- A.4 Scalability and Construction Cost: Graph construction takes less than 6 minutes and fewer than 230K tokens per user on average.HERO updates the graph incrementally without requiring global recomputation.