memory update rules

This work addresses the high latency, resource contention, and operational overhead caused by frequent state updates in streaming machine learning. The authors propose a probabilistic sparsification strategy that decouples inference from state persistence: while all events contribute to inference scoring, only those deemed highly informative trigger persistence. This approach enables precise control over the persistence path without requiring high-frequency in-memory control planes or cross-node coordination, while preserving unbiasedness of time-aggregated statistics. By integrating approximate statistics from disk-based key-value stores with variance-aware temporal aggregation modeling, the method reduces persistence events by up to 90%, substantially lowering I/O and serialization costs while maintaining or even improving downstream task performance.

This work addresses the challenges of real-world streaming sequential memory, which suffers from irregular sampling and non-stationary dynamics. Conventional vector symbolic architectures struggle with redundant, conflicting, and outdated information due to their simplistic accumulation mechanisms. To overcome these limitations, the authors propose the Sequential Relevance Memory Unit (SRMU), which integrates a time-decay function and a domain-agnostic relevance gating mechanism into the vector symbolic architecture. SRMU dynamically filters out irrelevant inputs prior to memory writing, enabling source-controlled memory updates without relying on post-hoc cleanup strategies. Evaluated on streaming state-tracking tasks, this approach improves memory-state alignment by 12.6% in similarity while reducing cumulative memory magnitude by 53.5%, thereby significantly enhancing both memory stability and fidelity to the true underlying state.

This work addresses the fundamental limitation of conventional neural networks, which rely on repeated data access and epoch-based training, thereby losing long-term consistency in irreversible data streams and degenerating into reactive filters devoid of temporal memory. To overcome this, the paper introduces Streaming Neural Networks (StNN), equipped with a stream-native execution algorithm (SNA) and streaming neurons that maintain persistent temporal states, enabling continuous, epoch-free learning. The study establishes the first minimal computational foundation for neural processing of irreversible streaming data, theoretically proving that the network’s state dynamics exhibit boundedness and contractivity. Through phase-space analysis and temporal modeling, the authors further demonstrate the stability and effectiveness of StNN in long-horizon stream processing tasks.

This work addresses the lack of efficient and interpretable memory mechanisms in sequence modeling by proposing the “HiPPO Zoo,” a unified framework that extends the HiPPO theory to construct five explicit memory architectures. The approach embeds key capabilities of modern state space models (SSMs)—such as online adaptive memory allocation and associative recall—into interpretable structures based on orthogonal polynomials. By integrating structured linear differential equations with streaming-compatible training, the method enables efficient online updates while preserving theoretical transparency. Experimental results demonstrate that the proposed models effectively replicate the performance of contemporary SSMs across multiple synthetic tasks, all while maintaining clear interpretability and controllability of the memory dynamics.

Traditional recurrent neural networks (RNNs) struggle to model long-range dependencies due to their fixed update mechanism, often suffering from memory decay. This work proposes the Selective Update RNN (suRNN), which introduces neuron-level binary gates that update hidden states only at information-rich timesteps, thereby decoupling recurrent dynamics from sequence length. For the first time, this approach enables neurons to adaptively learn when to update over time, combining sparse activation with nonlinear recurrent structure to construct gradient shortcut paths across timesteps while preserving RNN efficiency. Experiments demonstrate that suRNN matches or exceeds Transformer accuracy on benchmarks such as the Long Range Arena and WikiText, while significantly reducing computational and memory costs for long sequences.

This work addresses the long-standing challenge of credit assignment over extended sequences in streaming reinforcement learning under partial observability, where truncated backpropagation through time (BPTT) severely limits gradient horizons. The authors propose a novel streaming reinforcement learning method based on Recurrent Trace Units (RTUs), which for the first time enables exact real-time recurrent learning (RTRL) without replay buffers or batch updates. Leveraging a diagonal recurrent architecture, the approach incurs only linear time and memory complexity. It is compatible with both discrete and continuous control and supports online policy optimization. Empirical results demonstrate that the method significantly outperforms existing streaming baselines on MemoryChain, POPGym, and partially observable MuJoCo tasks, while closely approaching the performance of batch-based PPO.

Existing concept drift methodologies struggle to handle the complex, multidimensional, and non-stationary dynamics inherent in data streams encountered by autonomous learning systems. This work proposes a unified three-dimensional classification framework that systematically characterizes drift phenomena across temporal, data, and model streams, thereby integrating paradigms such as drift adaptation, continual learning, and temporal generalization. Through a comprehensive systematic review of 193 studies, the paper introduces and formally delineates novel concepts—including representational drift, semantic drift, and policy instability—exposes critical limitations of current approaches, identifies fundamental open challenges, and offers a clear roadmap toward building intelligent learning systems capable of sustainable, autonomous evolution.

This work addresses the instability and uncontrolled update magnitudes in streaming reinforcement learning caused by fixed step sizes. Inspired by the normalized least mean squares (NLMS) algorithm, we propose a novel approach that adaptively determines step sizes by inversely reasoning from desired update effects. Extending this principle to deep reinforcement learning, we introduce Intentional TD and Intentional Policy Gradient algorithms, which incorporate eligibility traces and diagonal scaling to enable stable and efficient learning in a purely streaming setting—without relying on experience replay buffers. Empirical results demonstrate that our method achieves state-of-the-art streaming performance across multiple benchmark tasks, matching or surpassing batch-based algorithms that depend on experience replay.

This work addresses the limitations of existing memory systems in large language models during sequential learning, where blindly accepting all local memory updates often leads to overwriting useful knowledge, over-specialization of rules, or recency bias. To mitigate these issues, the authors propose Janus, a plug-and-play memory controller that intelligently decides whether to accept candidate memory updates without altering the underlying update mechanism. Janus introduces a Memory Momentum Trigger to detect anomalous memory trajectories and leverages a lightweight hybrid evaluation set—constructed from coverage, boundary, and new-task samples—to efficiently compare the quality of old and new memories without replaying historical data. Experiments demonstrate that Janus consistently improves accuracy by 2.7–4.6 percentage points across six datasets, two large language models, and two memory updaters.