检索记忆

Streaming video reasoning requires models to operate in a setting where history grows without bound while meaningful evidence remains scarce. In such a landscape, relevant signal is like an oasis -- small, critical, and easily lost in a desert of redundancy. Enlarging memory only widens the desert; aggressive compression dries up the oasis. The real difficulty lies in discovering where to look, not how much to remember. We therefore introduce OASIS, a novel framework for streaming video reasoning that tackles this challenge through structured, on-demand retrieval. It organizes streaming history into hierarchical events and performs reasoning as controlled refinement -- short-context inference first, followed by semantically grounded retrieval only when uncertainty arises. As the retrieval is driven by high-level intent rather than embedding similarity, the retrieved memory is substantially more accurate and less noisy. Additionally, the mechanism is plug-and-play, training-free, and readily attaches to different streaming MLLM backbones. Experiments across multiple benchmarks and backbones show that OASIS achieves strong gains in long-horizon accuracy and compositional reasoning with bounded token cost and low request delay.

Existing visual trackers mainly operate in a non-interactive, fire-and-forget manner, making them impractical for real-world scenarios that require human-in-the-loop adaptation. To overcome this limitation, we introduce Interactive Tracking, a new paradigm that allows users to guide the tracker at any time using natural language commands. To support research in this direction, we make three main contributions. First, we present InteractTrack, the first large-scale benchmark for interactive tracking, containing 150 videos with dense bounding box annotations and timestamped language instructions. Second, we propose a comprehensive evaluation protocol and evaluate 25 representative trackers, showing that state-of-the-art methods fail in interactive scenarios--strong performance on conventional benchmarks does not transfer. Third, we introduce Interactive Memory-Augmented Tracking (IMAT), a new baseline that employs a dynamic memory mechanism to learn from user feedback and update tracking behavior accordingly. Our benchmark, protocol, and baseline establish a foundation for developing more intelligent, adaptive, and collaborative tracking systems, bridging the gap between automated perception and human guidance. The full benchmark, tracking results, and analysis are available at \href https://github.com/NorahGreen/InteractTrack.git this URL .

Recent advances in video large language models have demonstrated strong capabilities in understanding short clips. However, scaling them to hours- or days-long videos remains highly challenging due to limited context capacity and the loss of critical visual details during abstraction. Existing memory-augmented methods mitigate this by leveraging textual summaries of video segments, yet they heavily rely on text and fail to utilize visual evidence when reasoning over complex scenes. Moreover, retrieving from fixed temporal scales further limits their flexibility in capturing events that span variable durations. To address this, we introduce WorldMM, a novel multimodal memory agent that constructs and retrieves from multiple complementary memories, encompassing both textual and visual representations. WorldMM comprises three types of memory: episodic memory indexes factual events across multiple temporal scales, semantic memory continuously updates high-level conceptual knowledge, and visual memory preserves detailed information about scenes. During inference, an adaptive retrieval agent iteratively selects the most relevant memory source and leverages multiple temporal granularities based on the query, continuing until it determines that sufficient information has been gathered. WorldMM significantly outperforms existing baselines across five long video question-answering benchmarks, achieving an average 8.4% performance gain over previous state-of-the-art methods, showing its effectiveness on long video reasoning.

The Segment Anything Model 2 (SAM2) has emerged as a foundation model for universal segmentation. Owing to its generalizable visual representations, SAM2 has been successfully applied to various downstream tasks. However, extending SAM2 to the RGB-D video salient object detection (RGB-D VSOD) task encounters three challenges including limited spatial modeling of linear LoRA, insufficient employment of SAM's multi-scale features, and dependence of initialization on explicit prompts. To address the issues, we present Multi-Modal Mixture-of-Experts with Memory-Augmented SAM (M4-SAM), which equips SAM2 with modality-related PEFT, hierarchical feature fusion, and prompt-free memory initialization. Firstly, we inject Modality-Aware MoE-LoRA, which employs convolutional experts to encode local spatial priors and introduces a modality dispatcher for efficient multi-modal fine-tuning, into SAM2's encoder. Secondly, we deploy Gated Multi-Level Feature Fusion, which hierarchically aggregates multi-scale encoder features with an adaptive gating mechanism, to balance spatial details and semantic context. Finally, to conduct zero-shot VSOD without manual prompts, we utilize a Pseudo-Guided Initialization, where a coarse mask is regarded as a pseudo prior and used to bootstrap the memory bank. Extensive experiments demonstrate that M4-SAM achieves the state-of-the-art performance across all evaluation metrics on three public RGB-D VSOD datasets.

Unsupervised Anomaly Detection (UAD) aims to identify abnormal regions by establishing correspondences between test images and normal templates. Existing methods primarily rely on image reconstruction or template retrieval but face a fundamental challenge: matching between test images and normal templates inevitably introduces noise due to intra-class variations, imperfect correspondences, and limited templates. Observing that Retrieval-Augmented Generation (RAG) leverages retrieved samples directly in the generation process, we reinterpret UAD through this lens and introduce RAID, a retrieval-augmented UAD framework designed for noise-resilient anomaly detection and localization. Unlike standard RAG that enriches context or knowledge, we focus on using retrieved normal samples to guide noise suppression in anomaly map generation. RAID retrieves class-, semantic-, and instance-level representations from a hierarchical vector database, forming a coarse-to-fine pipeline. A matching cost volume correlates the input with retrieved exemplars, followed by a guided Mixture-of-Experts (MoE) network that leverages the retrieved samples to adaptively suppress matching noise and produce fine-grained anomaly maps. RAID achieves state-of-the-art performance across full-shot, few-shot, and multi-dataset settings on MVTec, VisA, MPDD, and BTAD benchmarks.

Vision-Language-Action (VLA) models have shown great performance in robotic manipulation by mapping visual observations and language instructions directly to actions. However, they remain brittle under distribution shifts: when test scenarios change, VLAs often reproduce memorized trajectories instead of adapting to the updated scene, which is a failure mode we refer to as the "Memory Trap". This limitation stems from the end-to-end design, which lacks explicit 3D spatial reasoning and prevents reliable identification of actionable regions in unfamiliar environments. To compensate for this missing spatial understanding, 3D Spatial Affordance Fields (SAFs) can provide a geometric representation that highlights where interactions are physically feasible, offering explicit cues about regions the robot should approach or avoid. We therefore introduce Affordance Field Intervention (AFI), a lightweight hybrid framework that uses SAFs as an on-demand plug-in to guide VLA behavior. Our system detects memory traps through proprioception, repositions the robot to recent high-affordance regions, and proposes affordance-driven waypoints that anchor VLA-generated actions. A SAF-based scorer then selects trajectories with the highest cumulative affordance. Extensive experiments demonstrate that our method achieves an average improvement of 23.5% across different VLA backbones (pi-0 and pi-0.5) under out-of-distribution scenarios on real-world robotic platforms, and 20.2% on the LIBERO-Pro benchmark, validating its effectiveness in enhancing VLA robustness to distribution shifts. Project page: https://vla-afi.github.io/.

Existing autonomous driving systems rely on onboard sensors (cameras, LiDAR, IMU, etc) for environmental perception. However, this paradigm is limited by the drive-time perception horizon and often fails under limited view scope, occlusion or extreme conditions such as darkness and rain. In contrast, human drivers are able to recall road structure even under poor visibility. To endow models with this recall ability, we propose the spatial retrieval paradigm, introducing offline retrieved geographic images as an additional input. These images are easy to obtain from offline caches (e.g., Google Maps or stored autonomous driving datasets) without requiring additional sensors, making it a plug-and-play extension for existing AD tasks. For experiments, we first extend the nuScenes dataset with geographic images retrieved via Google Maps APIs and align the new data with ego-vehicle trajectories. We establish baselines across five core autonomous driving tasks: object detection, online mapping, occupancy prediction, end-to-end planning, and generative world modeling. Extensive experiments show that the extended modality could enhance the performance of certain tasks. We will open-source dataset curation code, data, and benchmarks for further study of this new autonomous driving paradigm.

Embodied Question Answering (EQA) requires agents to navigate 3D environments, accumulate visual evidence, and reason over partial observations to answer questions. However, current agents struggle to maintain coherent, long-horizon behavior: planning remains reactive, causing inconsistent actions, while monolithic memories entangle all observations, hindering retrieval of the sparse but crucial evidence. We address these issues by reframing EQA through the lens of predictive processing, in which coherent behavior emerges from a prediction-correction loop grounded in stable priors. Guided by this perspective, we propose Predict Before You Explore (Pred-EQA), an architecture that integrates predictive planning with specialized memory. A high-level planner predicts where question-relevant evidence is likely to appear and generates a compact set of actionable exploration branches encoding long-horizon intent. A low-level executor then reduces uncertainty within these branches, revising predictions when they fail. A dual-memory system complements this process by separating slowly evolving structural priors from compact, question-relevant visual evidence, enabling consistent planning and efficient evidence accumulation. Through this prediction-guided exploration, Pred-EQA achieves coherent trajectories under partial observability. Experiments on OpenEQA and Express-Bench show that Pred-EQA achieves state-of-the-art results in both accuracy and exploration efficiency, demonstrating the benefits of prediction-driven embodied reasoning.

Multimodal biomedical Vision-Language Models (VLMs) exhibit immense potential in the field of Continual Learning (CL). However, they confront a core dilemma: how to preserve fine-grained intra-modality features while bridging the significant domain gap across different modalities. To address this challenge, we propose a comprehensive framework. Leveraging our 18-million multimodal and comprehensive medical retrieval database derived from PubMed scientific papers, we pioneer the integration of Retrieval-Augmented Generation (RAG) into CL. Specifically, we employ a multi-modal, multi-layer RAG system that provides real-time guidance for model fine-tuning through dynamic, on-demand knowledge retrieval. Building upon this, we introduce a dynamic knowledge distillation framework. This framework precisely resolves the aforementioned core dilemma by dynamically modulating the importance of the parameter space, the granularity of the distilled knowledge, and the data distribution of the reference dataset in accordance with the required level of detail. To thoroughly validate the clinical value of our strategy, we have designed a more rigorous Medical Generalist Task Incremental Learning (MGTIL) benchmark. This benchmark is engineered to simultaneously evaluate the model's capacity for adaptation to significant domain shifts, retention of subtle intra-domain features, and real-time learning of novel and complex medical tasks. Our method achieves stable SOTA results across all metrics, with https://github.com/CZZZZZZZZZZZZZZZZZ/PRIMED code available.

Existing retrieval-augmented approaches for Dense Video Captioning (DVC) often fail to achieve accurate temporal segmentation aligned with true event boundaries, as they rely on heuristic strategies that overlook ground truth event boundaries.The proposed framework, STaRC, overcomes this limitation by supervising frame-level saliency through a highlight detection module. Note that the highlight detection module is trained on binary labels derived directly from DVC ground truth annotations without the need for additional annotation.We also propose to utilize the saliency scoresas a unified temporal signal that drives retrieval via saliency-guided segmentation and informs caption generation through explicit Saliency Prompts injected into the decoder.By enforcing saliency-constrained segmentation, our method produces temporally coherent segments that align closely with actual event transitions, leading to more accurate retrieval and contextually grounded caption generation. We conduct comprehensive evaluations on the YouCook2 and ViTT benchmarks, where STaRC achieves state-of-the-art performance across most of the metrics.

Large Vision-Language Models (VLMs) have achieved remarkable success in multi-modal reasoning, but their inference time efficiency remains a significant challenge due to the memory overhead during decoding, especially when the query and answer of VLMs consist of long sequences of visual and text tokens. This paper presents AttentionPack, an adaptive and attention-aware optimization framework tailored for large vision-language models with improving memory-efficiency during decoding, focusing on addressing the challenges due to the increased high number of visual inputs and interactions, particularly in long-context tasks with multiple high-resolution images or videos. AttentionPack is novel in two aspects: (i) We introduce a multi-head attention compaction method for economically storing key and value matrices by exploiting the implicit low-rank structure, and (ii) we develop a token-specific attention-aware decompression mechanism to reduce latency overhead. Experimental results on multiple benchmarks demonstrate that AttentionPack improves memory efficiency by up to 8x, enabling higher batch sizes and faster batch inference while preserving the model output quality or longer context lengths for superior retrieval performance. We also report the effectiveness of AttentionPack combined with eviction, quantization and kernel fusion, showing further efficiency gains for resource-limited environments.

Vehicle trajectory prediction is critical for safe and efficient autonomous driving. However, its generalization and scalability are hindered by heavy reliance on real-time, online priors. To break this bottleneck, we introduce RAG-TP, a framework reframing the problem from relying on uncertain online perception to retrieving from a large-scale, structured knowledge base. RAG-TP enhances inference-time predictions by dynamically querying a heterogeneous knowledge base rich with scene topologies and motion patterns, using retrieved historical experiences as priors. We further design a dynamic fusion module based on a novel Retrieval-Driven Mixture-of-Experts (MoE). Unlike conventional parametric designs, this mechanism dynamically treats retrieved knowledge units as experts, weighting and integrating them via cross-attention to generate a dense context for final multi-modal trajectory decoding. By decoupling online inference from offline knowledge, this approach grounds predictions in a vast structured database, mitigating model hallucination, compensating for unreliable priors, and significantly enhancing robustness and domain adaptation. Extensive experiments show RAG-TP achieves excellent performance in map-based and map-free settings, demonstrating highly competitive results against specialized map-free methods while performing on par with state-of-the-art (SOTA) map-based models. It demonstrates significant advantages, particularly in cross-domain and zero-shot generalization. Our work provides a promising technical pathway toward building scalable and robust prediction systems for autonomous driving.

MLLMs exhibit strong reasoning on isolated queries, yet they operate de novo--solving each problem independently and often repeating the same mistakes. Existing memory-augmented agents mainly store past trajectories for reuse. However, trajectory-based memory suffers from brevity bias, gradually losing essential domain knowledge. More critically, even in truly multimodal problem-solving settings, it records only a **single-modality** trace of past behavior, failing to preserve how visual attention and logical reasoning jointly contributed to the solution. This is fundamentally misaligned with human cognition: semantic memory is both **multimodal and integrated**, preserving visual and abstract knowledge through coordinated but distinct representational streams. We thus introduce **ViLoMem**, a dual-stream memory framework that constructs compact, schema-based memory. It separately encodes visual distraction patterns and logical reasoning errors, enabling MLLMs to learn from their successful and failed experiences. Following a grow-and-refine principle, the system incrementally accumulates and updates multimodal semantic knowledge--preserving stable, generalizable strategies while avoiding catastrophic forgetting. Across nine multimodal benchmarks, **ViLoMem** consistently improves pass@1 accuracy and substantially reduces repeated visual and logical errors. Ablations confirm the necessity of dual-stream memory with explicit distraction-hallucination separation, demonstrating the value of error-aware multimodal memory for lifelong and cross-domain agentic learning.

Hierarchical Vision-Language-Action (VLA) models have rapidly become a dominant paradigm for robotic manipulation. It typically comprising a Vision-Language backbone for perception and understanding, together with a generative policy for action generation. However, its performance is increasingly bottlenecked by the action generation proceess. (i) Low inference efficiency. A pronounced distributional gap between isotropic noise priors and target action distributions, which increases denoising steps and the incidence of infeasible samples. (ii) Poor robustness. Existing policies condition solely on the current observation, neglecting the constraint of history sequence and thus lacking awareness of task progress and temporal consistency. To address these issues, we introduce OptimusVLA, a dual-memory VLA framework with Global Prior Memory (GPM) and Local Consistency Memory (LCM). GPM replaces Gaussian noise with task-level priors retrieved from semantically similar trajectories, thereby shortening the generative path and reducing the umber of function evaluations (NFE). LCM dynamically models executed action sequence to infer task progress and injects a learned consistency constraint that enforces temporal coherence and smoothness of trajectory. Across three simulation benchmarks, OptimusVLA consistently outperforms strong baselines: it achieves 98.6% average success rate on LIBERO, improves over pi_0 by 13.5% on CALVIN, and attains 38% average success rate on RoboTwin 2.0 Hard. In Real-World evaluation, OptimusVLA ranks best on Generalization and Long-horizon suites, surpassing pi_0 by 42.9% and 52.4%, respectively, while delivering 2.9x inference speedup.

Recent advances in Multimodal Large Language Models have greatly improved visual understanding and reasoning, yet their quadratic attention and offline training protocols make them ill-suited for streaming settings where frames arrive sequentially and future observations are inaccessible. We diagnose a core limitation of current Video-LLMs, namely Time-Agnosticism, in which videos are treated as an unordered bag of evidence rather than a causally ordered sequence, yielding two failures in streams: temporal order ambiguity, in which the model cannot follow or reason over the correct chronological order, and past-current focus blindness where it fails to distinguish present observations from accumulated history. We present WeaveTime, a simple, efficient, and model-agnostic framework that first teaches order and then uses order. We introduce a lightweight Temporal Reconstruction objective--our Streaming Order Perception enhancement--that instills order-aware representations with minimal finetuning and no specialized streaming data. At inference, a Past-Current Dynamic Focus Cache performs uncertainty-triggered, coarse-to-fine retrieval, expanding history only when needed. Plugged into exsiting Video-LLM without architectural changes, WeaveTime delivers consistent gains on representative streaming benchmarks, improving accuracy while reducing latency. These results establish WeaveTime as a practical path toward time-aware stream Video-LLMs under strict online, time-causal constraints. Code and weights will be made publicly available.

Continuous-time neural networks provide adaptive dynamics, but rely on a single hidden state to encode both fast input fluctuations and longer-term context. This shared representation forces rapidly changing inputs to overwrite slower contextual signals, causing the model to lose past information as new observations arrive. In contrast, biological perceptual systems maintain stable behaviour under evolving sensory input by integrating ongoing signals with stored associative patterns rather than relying on a single evolving state.Motivated by this distinction, we study a simple coupling of Liquid Time-Constant Networks (LTCs) with a Modern Hopfield Network (MHN) that serves as a content-addressable memory. At each time step, the liquid state is projected into a query, the MHN retrieves a memory vector, and the two representations are concatenated before a readout layer. We analyse this coupling under standard norm and Lipschitz assumptions and show that the combined representation remains bounded. We further show that the retrieval map contracts gradients for parameters upstream of the memory query, which provides a mechanism for reducing curvature in the loss landscape.On public time-series benchmarks, the coupled LTC-MHN model improves mean accuracy by 2.3% over competitive recurrent and continuous-time baselines and reduces the estimated Hessian trace by about an order of magnitude relative to a standalone LTC encoder, with the largest gains on classification tasks and competitive performance on a regression task. Qualitative analyses of training curves, loss landscapes, and latent embeddings support the interpretation that Hopfield retrieval smooths optimization and encourages more compact, linearly separable class manifolds. Code will be released upon publication.

Large multimodal models (LMMs) have shown promising performance for various document recognition tasks. However, LMMs adopt implicit modeling, and the parameters lack interpretability. Inspired by recent advances in human memory and learning research, we propose an explicit multiscale prototype memory that augments document recognition models, explicitly modeling recurrent layout and stylistic patterns across different spatial resolutions. A Memory Retrieval Mechanism enables local regions to sparsely attend to a few prototypes (e.g., image borders, tilted text); the retrieved compositional factors are concatenated with visual features and passed to the decoder, providing explicit region-wise structural context. Prototype memory consolidation updates and stabilizes prototypes via attention-weighted exponential moving average (EMA) strategy, while sparsity and anti-collapse regularization promote selective activation. We further adopt hierarchical memory for multi-resolution encoding. The proposed DREAM module is a plug-and-play component, allowing seamless integration into various encoder-decoder architectures. We validate on two tasks including document recognition on public datasets and the self-built DreamDoc dataset, and handwriting recognition on the SCUT-HCCDoc and SCUT-EPT datasets. Experimental results show that the proposed method is effective.

RGB-Thermal (RGBT) tracking aims to achieve robust object localization across diverse environmental conditions by fusing visible and thermal infrared modalities. However, existing RGBT trackers rely solely on initial-frame visual information for target modeling, failing to adapt to appearance variations due to the absence of language guidance. Furthermore, current methods suffer from redundant search regions and heterogeneous modality gaps, causing background distraction. To address these issues, we first introduce textual descriptions into RGBT tracking benchmarks. This is accomplished through a pipeline that leverages Multi-modal Large Language Models (MLLMs) to automatically produce texual annotations. Afterwards, we propose RAGTrack, a novel Retrieval-Augmented Generation framework for robust RGBT tracking. To this end, we introduce a Multi-modal Transformer Encoder (MTE) for unified visual-language modeling. Then, we design an Adaptive Token Fusion (ATF) to select target-relevant tokens and perform channel exchanges based on cross-modal correlations, mitigating search redundancies and modality gaps. Finally, we propose a Context-aware Reasoning Module (CRM) to maintain a dynamic knowledge base and employ a Retrieval-Augmented Generation (RAG) to enable temporal linguistic reasoning for robust target modeling. Extensive experiments on four RGBT benchmarks demonstrate that our framework achieves state-of-the-art performance across various challenging scenarios. The source code is available at https://github.com/IdolLab/RAGTrack.