EVOLVE-MEM: A Self-Adaptive Hierarchical Memory Architecture for Next-Generation Agentic AI Systems

Rishi Ashish Shah, Ujjwal Kakar, Shashvat Singhal, Dinesh K Vishwakarma

Keywords: Agentic AI, Large Language Models, Memory Architecture, Hierarchical Systems, Self-Adaptive Systems

TL;DR: EVOLVE-MEM: a Hierarchical Self Evolving Memory Architecture for Agentic Systems integrating dynamic vector ingestion, multi-level clustering & summarization, and automated self-improvement.

Abstract: This paper introduces EVOLVE-MEM, a novel self-adaptive hierarchical memory framework designed to overcome the inherent limitations of fixed-size context windows and static memory architectures hindering long-term retention and adaptation in modern AI systems. The architecture is structured around three interconnected tiers: a Dynamic Memory Network leveraging dense semantic storage and embeddings for raw experience ingestion, labeled Level 0; a Hierarchical Memory Manager that organizes these embeddings into multi-level abstractions, utilizing adaptive clustering and large language models: Level 1 generates contextual summaries, and Level 2 extracts high-level principles; and a Self-Improvement Engine continuously monitoring key performance metrics: accuracy, retrieval latency, and coverage to autonomously trigger memory reorganization when thresholds are exceeded, ensuring the system evolves with changing data distributions. The system combines dynamic clustering with empirically tuned similarity thresholds, multi-level retrieval routing, and a robust answer-patching pipeline for post-processing raw LLM outputs to support complex temporal, causal, and multi-hop reasoning tasks with high fidelity and less generic responses. The framework achieves 58.3% overall accuracy, evaluated on the LoCoMo dataset, surpassing SOTA baselines across five reasoning categories. EVOLVE-MEM’s transparent retrieval path tracking, modular design for extensibility, and fully automated adaptation establish a new paradigm for truly agentic AI systems capable of sustained operation and continuous learning in complex, dynamic environments.

Archival Option: The authors of this submission want it to appear in the archival proceedings.

Submission Number: 62