adaptive memory fusion

This work addresses the challenge of jointly achieving domain generalization and continual learning in multi-task settings, where catastrophic forgetting commonly occurs. We propose MIRA, a unified framework that integrates Hopfield-style associative memory into adapter architectures. Its core innovation lies in a reconfigurable shared backbone coupled with sample-level dynamic retrieval, enabled by post-training key learning and affine composition-based updates for task-adaptive modulation. This design emulates neuromodulatory regulation of a single neural circuit, facilitating rapid task switching and persistent knowledge retention. On standard benchmarks, MIRA achieves state-of-the-art out-of-distribution accuracy for domain generalization while significantly outperforming dedicated continual learning methods. Notably, it demonstrates superior knowledge retention across incremental tasks, effectively mitigating catastrophic forgetting without compromising generalization.

This work proposes a lightweight closed-loop adaptive framework to address the stability-plasticity dilemma in online representation learning over non-stationary data streams, alongside the challenges of limited edge computing resources and concept drift. The core innovation introduces an endogenous residual feedback mechanism termed "shock ratio," which formulates representation learning as a closed-loop control process. By leveraging adaptive incremental gating and a continuous plasticity controller, the framework achieves linear computational complexity while dynamically interpolating between memory retention and rapid adaptation. Real-world evaluations across diverse smart city scenarios demonstrate that this approach enables efficient, forgetting-resistant real-time learning under strict resource constraints. It significantly reduces early warning lead times to 8–9 steps, accelerates post-drift recovery, and effectively enhances both anomaly recall rates and noise robustness.

This study addresses the challenge of rapid generalization under dynamic working memory maintenance and task switching. Inspired by the dorsal–ventral (DV) axis hierarchy in the hippocampus, we propose the GATE model—a biologically grounded neural architecture. Methodologically, GATE introduces a novel three-dimensional DV organization enabling graded representations from perceptual detail to abstract schema; it implements an EC3–CA1–EC5–EC3 re-entrant loop with dynamic gating in CA3 and EC5 to support selective memory maintenance and readout. Leveraging neurodynamical modeling and biologically constrained learning rules, the model replicates hallmark hippocampal neuron types—including splitter, trajectory, and delay-active cells. Empirically, GATE achieves zero-shot rapid generalization upon abrupt environmental, cue, or task changes, and its learned representations exhibit cross-task transferability. These results significantly advance the adaptability and generalization capacity of brain-inspired memory systems.

This work addresses the lack of a unified computational interpretation for neural policy gating mechanisms. We propose GateMod, a theoretically grounded gating framework that couples task structure with neural circuit dynamics via the principle of free-energy minimization. GateMod comprises two core components: GateFlow—a continuous-time energy-flow model—and GateNet—a soft-competitive recurrent network—enabling emergent gating for skill composition and behavioral planning. We formally prove GateMod’s global exponential convergence and robustness under perturbations. Empirically, GateMod achieves significant performance gains over state-of-the-art methods in multi-agent cooperative tasks and human multi-armed bandit experiments. Crucially, it provides the first quantitative demonstration of how task structure modulates gating behavior through neural energy dynamics. By offering a computationally precise and empirically testable account, GateMod establishes a principled theoretical foundation for understanding strategy selection in prefrontal–basal ganglia circuits.

Linear Transformers suffer from weak memory control and limited performance in long-context modeling and retrieval tasks. To address this, we propose Gated DeltaNet—a parallelizable hybrid architecture. Its core innovation is the first-ever gated delta update mechanism, enabling coordinated rapid memory erasure and precise incremental updates. Additionally, it seamlessly integrates sliding-window attention with a Mamba2-style linear state space model (SSM). This design jointly preserves local sensitivity and global contextual modeling capability, substantially improving training efficiency and inference stability. Experiments demonstrate that Gated DeltaNet consistently outperforms both Mamba2 and DeltaNet across diverse benchmarks—including language modeling, commonsense reasoning, context retrieval, length extrapolation, and long-text understanding—while achieving superior generalization and higher training throughput.

This work addresses the challenge of balancing stability and plasticity in continual learning under sequential data scenarios. Inspired by the modular organization of the human brain, the authors propose MoRe, a novel framework that constructs a theoretically identifiable hierarchical modular structure in representation space. MoRe decomposes knowledge into shared foundational modules and task-specific modules, enabling module reuse, alignment, and expansion. By leveraging temporal delayed dependencies to uncover intrinsic sequence structures and integrating modular learning with identifiability constraints, MoRe achieves structured knowledge organization and protection without requiring explicit task boundaries. Experiments on synthetic benchmarks and activation data from large language models demonstrate that MoRe learns interpretable hierarchical representations and significantly improves the stability-plasticity trade-off in continual learning.

研究探讨了状态空间模型(SSMs)中门控机制对记忆与上下文学习的影响，通过理论和实验揭示了其在训练动态及泛化中的关键作用。

为解决长期记忆中临时信息覆盖持久知识的问题，提出LifeFuse-Mem框架，通过生命周期感知的更新方法分离不同类型的信息。

This work addresses the limitations of existing reinforcement learning methods, which often fail to explicitly model the local geometric structure of the state space and struggle to effectively disentangle and adaptively integrate dynamic information with reward signals. Inspired by neuroscience, the paper proposes a novel framework that leverages Locally Linear Embedding (LLE) to capture the local manifold structure of states, combines it with standard reinforcement learning objectives to extract reward-relevant features, and introduces a cortex-inspired gating attention mechanism to adaptively fuse these two representations based on the current state. By integrating LLE with adaptive feature fusion for the first time, the approach emulates the brain’s mechanisms for information segregation and integration, enabling structured representation learning. Empirical results across multiple benchmark tasks demonstrate significant improvements in both sample efficiency and performance, validating the efficacy of modeling local state geometry and dynamically selecting features.

This study addresses the cross-session association blind spots, core memory forgetting, and topological stagnation of LLM agents by proposing an adaptive memory architecture grounded in the Complementary Learning Systems framework. Methodologically, it optimizes retrieval through immediate reflection and asynchronous consolidation. It introduces modulated PageRank, topological load decay, and a SUPERSEDES filtering mechanism to enable dynamic graph evolution, while integrating Hebbian plasticity, hybrid retrieval (vector/BM25/graph), Reciprocal Rank Fusion (RRF), and Directed Acyclic Graphs (DAGs) to enhance representational capacity. Evaluated on the LoCoMo benchmark, the proposed approach improves Recall@5 by 38.9%, eliminates hallucinations, and achieves 100% retention of long-term core memories.