Rethinking Memory as Continuously Evolving Connectivity

The evolution of memory systems in AI has transitioned from simple static storage to more structured and dynamic architectures. Early models like Memory Networks (Weston et al., 2014) and Neural Turing Machines (Graves et al., 2014) attempted to mimic human-like long-term memory but faced limitations in flexibility and scalability. Recent advances introduced hierarchical and graph-based memory structures (Han et al., 2022; Long et al., 2023), aiming to improve connection richness and adaptability. Despite these efforts, most existing systems rely on fixed, hand-crafted pipelines, which hinder their ability to dynamically reconfigure based on environmental feedback. The need for a memory mechanism that can self-organize, evolve, and consolidate knowledge over time remains unmet. This paper builds upon these foundations, proposing a self-evolving, graph-based memory system that continuously refines its topology through feedback, inspired by cognitive science principles of structural and functional plasticity.

Current memory-augmented models often treat memory as a static repository, which leads to brittle performance in dynamic environments. The fixed representations and retrieval pipelines cannot adapt to feedback, task variations, or heterogeneous signals, resulting in issues like under-connection, irrelevant retrieval, and inflexibility in updating memory units. These problems severely limit the ability of models to perform long-term, autonomous learning and reasoning. Moreover, existing methods lack mechanisms for long-term consolidation, preventing the formation of stable, structural memory regions. Addressing these issues requires a fundamentally new approach that models memory as a flexible, self-organizing network capable of continuous evolution based on environmental cues.

FluxMem introduces a three-stage memory evolution framework that models memory as a heterogeneous graph with semantic, episodic, and procedural layers. The first stage establishes initial connections by integrating relevance scores from dense embeddings, lexical matching, and LLM verification. The second stage employs a feedback loop, where environmental signals guide the addition or removal of links, refining the graph structure to better support current tasks. The third stage clusters successful trajectories to induce reusable skills, monitored by the PEMS metric, which quantifies the structural maturity of memory nodes. This approach enables continuous, self-guided adaptation of memory topology, overcoming the rigidity of static systems. The integration of relevance-based retrieval, feedback correction, trajectory clustering, and skill induction into a unified framework represents a significant innovation, providing a scalable and flexible solution for lifelong learning in AI agents.

• �� Initial Connection Formation: At each task step, the system retrieves relevant facts from the semantic layer by calculating a hybrid relevance score combining dense embedding similarity, lexical matching, and LLM verification. This establishes initial edges between current observations and stored knowledge. Simultaneously, episodic experiences are retrieved via embedding similarity, and applicable skills are inherited from related episodes through existing distillation links, forming a local subgraph.
• �� Feedback-Driven Refinement: During task execution, environmental feedback signals (success, failure, errors) are used to dynamically modify the graph. Missing links are added by identifying semantically proximate but unactivated nodes; irrelevant links are pruned to reduce noise. Internal node content is reshaped to align abstraction levels, either by expanding details or abstracting redundant information. These edits are iteratively applied until task success or a maximum refinement round is reached.
• �� Long-Term Consolidation: After task completion, trajectories are clustered based on semantic similarity. Skills are extracted via LLM induction and refined through iterative evaluation guided by PEMS. The stability and maturity of skills are monitored, and validated skills are stored as procedural nodes, enriching the memory for future tasks. Offline consolidation ensures the memory graph evolves into a stable, self-organized structure capable of supporting complex reasoning and multi-task learning.
• �� Connection and skill updates are guided by relevance scores, feedback signals, and PEMS, ensuring continuous adaptation and structural stability. The entire process is integrated with online real-time updates during task execution and offline long-term consolidation, forming a cohesive, self-improving memory system.

The experimental setup involves three benchmark datasets: LoCoMo for long-context reasoning, Mind2Web for web navigation, and GAIA for general assistant tasks. Baselines include static memory models like MemoryOS and Nemori, as well as evolving memory systems such as MemEvolve and Flash-Searcher. Evaluation metrics encompass LMJ scores, success rates, Action F1, and Element Accuracy. Hyperparameters like the number of refinement rounds T and convergence threshold ϵ are tuned to analyze their impact on performance. Ablation studies systematically remove each stage to assess their contribution, revealing that feedback-driven refinement (Stage II) is critical for accuracy, while long-term consolidation (Stage III) enhances performance in multi-step tasks. Results consistently show that FluxMem outperforms baselines, with significant improvements in reasoning accuracy, navigation success, and multi-task transferability. The experiments demonstrate the framework’s robustness, scalability, and adaptability across diverse scenarios.

In the LoCoMo benchmark, FluxMem achieves a LMJ score of 95.06, far exceeding the baseline of 81.23, illustrating its superior long-context reasoning ability. On Mind2Web, success rates improve from 52.12 to 73.6 in realistic settings, outperforming models like AWM and MemoryOS by margins of 15-20%. In GAIA, success rate increases from 52.12 to 73.6, surpassing MemEvolve and Flash-Searcher, indicating excellent cross-task transfer. Ablation results highlight that removing feedback refinement reduces LMJ scores by over 10%, confirming its importance. The number of refinement rounds T correlates positively with performance, with diminishing returns beyond T=3. The Procedure Evolution Maturity Score (PEMS) stabilizes after several rounds, indicating convergence of memory structure. Overall, these findings validate the effectiveness of the three-stage evolution process in building adaptive, self-organizing memory networks.

FluxMem can be directly applied to intelligent assistants, autonomous agents, and knowledge management systems requiring lifelong learning. Its ability to dynamically refine memory connections makes it suitable for complex reasoning, multi-step planning, and real-time decision-making in industries like healthcare, finance, and customer service. The framework supports continuous knowledge updates, enabling systems to adapt to new information without retraining from scratch. Long-term, it paves the way for fully autonomous AI agents capable of self-organizing knowledge bases, improving scalability and robustness in real-world applications. Additionally, its modular design allows integration with multi-modal data sources, broadening its applicability to robotics, multimedia analysis, and interactive AI systems.

Despite its promising results, FluxMem faces challenges such as high computational overhead due to iterative feedback and connection updates, which may hinder deployment in resource-constrained environments. Its reliance on static datasets limits validation in real-time streaming scenarios, where memory decay and continuous data flow are critical. Parameter sensitivity (e.g., T, ϵ) requires careful tuning, lacking adaptive mechanisms for diverse tasks. Moreover, the offline consolidation process may introduce latency, affecting real-time responsiveness. Future work should focus on optimizing efficiency, developing adaptive parameter tuning, and extending the framework to handle active memory decay and streaming data for truly lifelong learning systems.