Learnijoy
ResearchAI Engineering & DevToolsAI Research

ProGraph Memory Boosts LLM Agent Multi-Hop Reasoning

Shengtong Zhu· July 23, 2026 View original

Summary

This paper introduces ProGraph, a two-layer memory architecture for LLM agents that significantly improves multi-hop reasoning and precision recall. It uses narrative profile expansion and compression residuals to enable implicit cross-entity traversal without explicit knowledge graph construction.

Long-term memory is crucial for LLM agents that need to maintain context and interact across multiple sessions. However, existing memory benchmarks often focus on simple, single-hop recall, failing to adequately test an agent's ability to make multi-hop associations between pieces of information. Researchers have developed ProGraph (Profile-Graph Memory), a novel two-layer memory architecture. This system leverages "profile expansion," which uses substring matching of entity names within LLM-generated narratives to traverse connections implicitly, avoiding the overhead of building explicit knowledge graphs. It also incorporates "compression residuals" – exact dates, quantities, and named items extracted alongside profile updates at no additional API cost. ProGraph was evaluated on a new multi-hop memory benchmark called MemHop, achieving 80.1% accuracy, comparable to a full-context reference model. It also outperformed other RAG methods on precision recall benchmarks, demonstrating that profile expansion drives multi-hop reasoning while compression residuals enhance precision.

Why it matters

For professionals developing or deploying LLM agents, ProGraph offers a more robust and efficient memory solution, enabling agents to perform complex, multi-hop reasoning and maintain higher factual accuracy across extended interactions.

How to implement this in your domain

  1. 1Adopt the ProGraph memory architecture for LLM agents requiring advanced long-term memory and multi-hop reasoning capabilities.
  2. 2Integrate profile expansion mechanisms to allow agents to implicitly traverse connections between entities mentioned in their narratives.
  3. 3Implement compression residuals to co-extract precise facts (dates, quantities) during profile updates for improved recall.
  4. 4Utilize the MemHop benchmark to rigorously test and validate the multi-hop reasoning abilities of your LLM agents.
  5. 5Continuously refine the agent's narrative generation and profile update strategies to optimize memory performance.

Who benefits

AI/ML DevelopmentCustomer ServiceGamingEducationHealthcare

Key takeaways

  • ProGraph is a two-layer memory architecture for LLM agents, enhancing long-term memory.
  • It uses narrative profile expansion for implicit multi-hop reasoning without explicit knowledge graphs.
  • Compression residuals improve precision recall by co-extracting exact facts.
  • ProGraph significantly outperforms other RAG methods on multi-hop and precision benchmarks.

Original post by Shengtong Zhu

"arXiv:2607.19359v1 Announce Type: new Abstract: Long-term memory is essential for LLM agents that interact across sessions, yet current memory benchmarks primarily evaluate single-hop recall, leaving multi-hop association largely unmeasured. We make three contributions. First, we…"

View on X

Originally posted by Shengtong Zhu on X · view source

Want to go deeper?

Turn these trends into skills with Learnijoy's hands-on AI & tech courses.

Explore courses

More in AI Engineering & DevTools

AI ResearchAI Engineering & DevTools

New Adaptive Filter Improves Time-Series Prediction with Input Noise

Researchers developed the RFFBCGA algorithm, a new nonlinear adaptive filter that effectively mitigates both input and output noise in time-series prediction. This method maintains a fixed network structure while enhancing robustness across various noise scenarios.

Yi Peng, Haiquan Zhao, Jinhui HuJul 23, 2026
AI ResearchAI Engineering & DevTools

New Algorithm Learns Local Causal Structures with Latent Variables

Researchers propose LoCaLS, a new algorithm for learning local causal structures around a target variable from observational data, even when latent variables and selection bias are present. LoCaLS achieves high accuracy with significantly less computational effort than global causal discovery methods.

Zheng Li, Hao Zhang, Ruxin Wang, Ruichu Cai, Kun Zhang, Feng XieJul 23, 2026
AI Engineering & DevToolsAI Research

New Framework Evaluates AI Robustness with Minimum-Norm Attacks

Researchers introduce a unified framework for evaluating adversarial robustness using a comprehensive pool of minimum-norm attacks and robustness-perturbation curves across multiple norms. This approach addresses limitations of fixed-epsilon evaluations, providing a more stable and controllable assessment of AI model defenses.

Luca Scionis, Luca Melis, Maura Pintor, Fabio Brau, Ambra Demontis, Giorgio Fumera, Fabio Roli, Battista BiggioJul 23, 2026