AuthorsYuanhua Lin, Yile Li, Zhiyuan Zhao et al.
TL;DR
HERO uses a human-profile guided heterogeneous memory graph with iterative cue activation to reach 56.06% F1 and 87.99% accuracy on LoCoMo.
Read our summary here, or open the publisher PDF on the next tab.
THE PROBLEM
HERO targets two issues: information loss from compression and semantic drift from rewriting, which discard fine-grained but later useful details.
When long-term agents rely on LLM summaries or profiles, they miss subtle temporal references and preference changes, degrading factual QA and personalized reasoning.
HOW IT WORKS
HERO’s core mechanism builds a heterogeneous memory graph with Episodic Traces, Episodic Units, Episodic Cues, Profile Insights, and Profile Cues plus Personalized PageRank retrieval.
You can think of HERO like a hippocampus-backed card catalog: raw dialogues stay as full episodes, while compact human profiles act as navigational signposts through the graph.
This design lets HERO follow multi-hop cue paths and profile-guided links across sessions, something a plain context window or vanilla RAG cannot support reliably.
DIAGRAM
This diagram shows how HERO runs profile-aware cue activation and graph traversal over the heterogeneous memory graph to retrieve evidence for a query.
DIAGRAM
This diagram shows how HERO is evaluated on LoCoMo and PERSONAMEM, including ablations without profiles and cue expansion.
PROCESS
Memory Construction
HERO segments dialogue into Episodic Traces and Episodic Units, then extracts Episodic Cues plus Profile Insights and Profile Cues to build the heterogeneous graph.
Initial Anchor Identification
HERO encodes the query, extracts entities Eq, and matches them to cue nodes in the graph to form the initial active set S0.
Iterative Context Expansion
HERO repeatedly follows cue to Episodic Units and Profile Insights, filters by query relevance, and activates new Episodic Cues and Profile Cues.
Global Relevance Propagation and Evidence Ranking
HERO combines cue activations with dense similarity, runs Personalized PageRank over Episodic Traces, and returns top k raw traces to the LLM.
KEY CONTRIBUTIONS
Human-profile enhanced retrieval optimization framework
HERO introduces a human-cognition-inspired framework where Profile Insights and Profile Cues guide retrieval instead of replacing raw memories, enabling profile-aware reasoning.
Heterogeneous memory graph with episodic evidence
HERO designs a graph over Episodic Traces, Episodic Units, and Episodic Cues that preserves complete dialogue text as final evidence for faithful reasoning.
Profile-aware cue activation and graph retrieval mechanism
HERO’s iterative cue activation plus Personalized PageRank retrieves explicit facts and implicit cross-episodic dependencies, achieving 56.06% F1 on LoCoMo and 70.63% accuracy on PERSONAMEM.
RESULTS
Overall ACC
87.99%
+3.96 over EverMemOS
Overall F1
56.06%
+10.58 over EverMemOS
Overall BLEU-1
48.38%
+11.44 over EverMemOS
PERSONAMEM Overall Avg
70.63%
+4.25 over EverMemOS
On LoCoMo, which tests single-hop, multi-hop, temporal, and open-domain QA over long conversations, HERO’s 87.99% accuracy and 56.06% F1 show strong factual grounding. On PERSONAMEM, HERO’s 70.63% overall accuracy demonstrates robust personalized reasoning over evolving user traits and preferences.
BENCHMARK
On LoCoMo, which tests single-hop, multi-hop, temporal, and open-domain QA over long conversations, HERO’s 87.99% accuracy and 56.06% F1 show strong factual grounding. On PERSONAMEM, HERO’s 70.63% overall accuracy demonstrates robust personalized reasoning over evolving user traits and preferences.
BENCHMARK
Overall F1 on LoCoMo across HERO and strong long-term memory baselines.
KEY INSIGHT
HERO reaches 56.06% F1 on LoCoMo while using only 248 tokens on average and keeping retrieval latency under one second.
This is surprising because many assume better long-term memory requires larger prompts and heavier online LLM reasoning, but HERO shifts complexity into offline graph construction instead.
WHY IT MATTERS
HERO unlocks faithful, profile-aware long-term memory where agents reason over raw episodic text guided by human profiles rather than lossy summaries.
Builders can now create agents that track multi-session facts and evolving preferences over tens of thousands of tokens without bloated prompts or brittle one-shot summaries.
Peilin Feng, Suorong Yang, Soujanya Poria
arXiv 2026 · 2026
Σ-Mem maintains historical competence evidence matrices M_p, a peer relationship matrix G, a memory readout r_{p,t}, and residual steering δ_{p,t} to guide multi-agent trust decisions. On the mixed counterfactual benchmark, Σ-Mem raises Qwen3-0.6B accuracy at CF@90 from 46.22% with the base model to 71.10% with Σ w/ G.
Nikhil Verma
· 2026
Focus Agent adds start_focus, complete_focus, a persistent Knowledge block, and an optimized Persistent Bash plus String-Replace Editor scaffold to actively compress context during long software-engineering tasks. On five hard SWE-bench Lite instances against a Baseline ReAct agent, Focus Agent achieves 22.7% token reduction (14.9M → 11.5M) while matching 3/5 = 60% task success.
Xiaohui Zhang, Zequn Sun et al.
· 2026
ActMem transforms dialogue history into atomic facts via Memory Fact Extraction, groups them with Fact Clustering, links them through a Memory KG Construction module, and uses Counterfactual-based Retrieval and Reasoning for action-aware answers. On ActMemEval, ActMem reaches 76.52% QA accuracy with DeepSeek-V3, beating LightMem’s 63.97% by 12.55 points and NaiveRAG’s 61.54%.
Answers use this explainer on Memory Papers.
Checking…