From Static Recommendations to Autonomous Decisions: A Framework for Agentic Recommendation Systems with Memory, Reasoning and Tool Use
International Journal of Sustainability and Innovation in Engineering · 2026
DOI: 10.56830/ijsie202606 ↗Kurz gefasst
Fachartikel in der Zeitschrift International Journal of Sustainability and Innovation in Engineering, 2026. Verfasst von Tirth Lakhani Lakhani. Die Arbeit nennt 65 Literaturangaben.
Abstract (Originalsprache)
Agentic recommendation systems can go beyond static rankings by grasping the intent behind actions, respecting constraints, calling for all necessary tools, communicating tradeoffs, and learning from feedback. However, controlled evidence on the cumulative contribution of these capabilities remains limited. This study created a five-agent system that includes a user-preference agent, product-retrieval agent, recommendation agent, explanation agent, and feedback-learning agent, coordinated by a deterministic orchestrator. The study conducted a comparative offline experiment using MovieLens 25M and Amazon reviews’23 electronics data, with 20,000 users, 20,000 ranking tasks, 300,000 sequential interactions, 10,000 feedback episodes, and 900 explanation judgments. The study measured popularity ranking, SASRec, single-agent LLM-RAG, and the proposed framework using NDCG@10, Recall@10, Hit Rate@10, MRR@10, constraint satisfaction, catalog coverage, diversity, novelty, explanation quality, reliability, latency, cost, and fallback behavior. The framework provided broader support and demonstrated traceability and auditability of memory, anchoring, reasoning, tool use, and recovery from failure. There was no model with an overall statistically significant advantage after Holm correction (adjusted p < .05). The available analytical outputs did not support reporting Krippendorff’s alpha, drift-recovery rate, incorrect-memory-update rate, post-feedback NDCG@10 changes, and operational threshold rates, which restricted claims of evaluator agreement, adaptation magnitude, and threshold compliance. The results show that absolute ranking accuracy is not the only metric for assessing an agentic recommendation system. Future studies should focus on live A/B testing at scale, dynamic catalogs, privacy-preserving memory, real feedback, fairness analysis for subgroups, and deployment costs.
Zitieren
APA
Lakhani, T. L. (2026). From Static Recommendations to Autonomous Decisions: A Framework for Agentic Recommendation Systems with Memory, Reasoning and Tool Use. International Journal of Sustainability and Innovation in Engineering. https://doi.org/10.56830/ijsie202606
Vancouver
Lakhani TL. From Static Recommendations to Autonomous Decisions: A Framework for Agentic Recommendation Systems with Memory, Reasoning and Tool Use. International Journal of Sustainability and Innovation in Engineering. 2026. doi:10.56830/ijsie202606
BibTeX
@article{Lakhani2026,
author = {Lakhani, Tirth Lakhani},
title = {From Static Recommendations to Autonomous Decisions: A Framework for Agentic Recommendation Systems with Memory, Reasoning and Tool Use},
journal = {International Journal of Sustainability and Innovation in Engineering},
year = {2026},
doi = {10.56830/ijsie202606}
}
EUVIMED