Memory-Augmented VLMs

Large language models (LLMs) and vision-language models (VLMs) are usually frozen after deployment, so they do not learn from the cases they solve. This is especially concerning in medicine, where new clinical evidence, updated guidelines, and new therapies can change established practice. Fine-tuning can update the model, but it requires access to model weights and additional training. Parameter-free methods avoid training, but they may overfit a fixed validation set, lack reliable domain knowledge, or lose visual details by saving experience only as text. To address these limitations, we present a model-agnostic framework that allows frozen LLMs and VLMs to learn from deployment experience through three forms of external expertise: a Skill that guides reasoning and tool use, a Knowledge Memory that stores reliable facts supported by earlier cases or trusted external evidence, and a Multimodal Knowledge Base that keeps visual examples and guides the model to relate each retrieved case to the current image. Instead of relying on a fixed validation set, a validation strategy keeps an update only if it helps on new cases without degrading performance on earlier ones. Across six benchmarks covering clinical diagnosis, clinical workflows, medical reasoning, and medical and non-medical visual reasoning, and with four open-weight and closed-source base models, our framework improves performance during online deployment by up to 34.2% over the base model on medical tasks, generalizes to unseen cases, transfers to other models without further optimization, and works in non-medical domains.

Multimodal assistants answer questions from long-term memories that contain images. After retrieval, each retrieved image reaches the answering model either as pixels, at about a thousand visual tokens per image, or as a stored text proxy that often misses the detail the question asks about. We find that the benefit of pixels usually comes from one or two retrieved memories, and that it can be predicted before the answering model runs, without reading any full-resolution image. In PixelTriage, a plug-in placed after retrieval, a small model that does not generate text reads the dialogue, a short note and a thumbnail of each retrieved memory and predicts how much its pixels would add. It is trained on synthetic memory episodes labeled by a frozen 27B model that answers each question with and without each memory's pixels. With a 7B answering model, PixelTriage lies on the accuracy--cost frontier of M$^3$Exam, DMV and MemEye and uses 11--23% of the visual tokens without a significant loss of accuracy. On DMV it answers 2.9 times faster than opening all images. It outperforms retrieval order and uniform down-sizing at equal budgets and transfers to other memory systems and to a 397B answering model.

When searching for an object, people choose their next move by considering both likely target locations and places already explored. The current view can cue place-associated memories, bringing target relevance and prior exploration into the same spatial context. In many zero-shot object navigation (ZSON) methods, however, vision-language models (VLMs) infer promising search areas from egocentric images, while exploration history is represented separately, e.g., as text or maps. This separation either requires an additional fusion step or leaves the correspondence between memory and route choices implicit for the VLM to recover. We instead make exploration memory directly visible on visual route choices. We propose MarvisNav, a ZSON framework that maintains a topological graph and projects candidate nodes together with their exploration states onto egocentric views as memory-bearing visual route choices. These states capture local exploration progress beyond binary visitation. By binding exploration state directly to each visual candidate, MarvisNav enables the VLM to jointly evaluate target relevance and exploration state without a separate post-hoc fusion or reranking stage. Without policy training, MarvisNav achieves state-of-the-art performance on HM3D (81.2% SR and 42.5% SPL), while remaining competitive on MP3D. It also outperforms representative VLM-based methods with far fewer VLM calls (e.g., 7.5% of WMNav). Real-robot experiments across diverse scenes further validate its practical deployability. Beyond MarvisNav, our study shows that memory representation shapes VLM decisions and ZSON performance, highlighting that effective memory use depends not only on its availability, but also on how it is represented. Code and project page will be available at https://wangjincheng1998.github.io/MarvisNav/.

Despite their impressive performance on a wide range of video understanding tasks, current Vision Language Models (VLMs) are predominantly designed for offline scenarios and struggle to handle online streaming videos that demand low latency response. Several studies have explored memory and token compression strategies in an attempt to adapt offline VLMs for streaming video understanding tasks. However, through our probing experiment, we identify that most existing works tend to progressively lose long context information as length of input stream increases. To address this, we propose ReMem, a novel training-free adaptation technique that enables VLMs to process streaming videos of arbitrary lengths while improving their long context information retention capability. ReMem exploits memory from two perspectives, implemented as two core components. The Streaming Context Memory (SCM) continuously compresses historical context with query-independent attention. The Retrieved Vision Memory (RVM) then retrieves the most salient, query-relevant context from memory to augment the VLM's input. Comprehensive experiments demonstrate that the proposed ReMem achieves state-of-the-art (SOTA) performance across a variety of widely used benchmarks, spanning both streaming video and general long video understanding tasks.

Most vision-language-action (VLA) models rely on current observations and lose task-relevant evidence once it leaves view, limiting performance on long-horizon, memory-dependent tasks. Existing efforts incorporate compressed historical features or sparse visual keyframes. However, isolated snapshots can leave the policy uncertain about what changed during past interactions and which action should follow. To overcome this limitation, we propose EvoMem-VLA, which constructs state-evolution memory by explicitly encoding and retaining observed changes between historical states. These change representations preserve evidence of interaction outcomes, allowing the policy to track task progress beyond isolated snapshots. Specifically, we introduce conditional delta tokenization to encode ordered frame pairs into directional, source-conditioned delta tokens, each associated with its corresponding state evidence. A shared VLM backbone supports task-adaptive routing: normal long-horizon tasks follow a direct action route, whereas multi-stage tasks use a subtask route that generates an executable subtask as an additional input for action generation. With a single jointly trained policy for each simulation benchmark, EvoMem-VLA achieves success rates of 80.7% on RMBench, 82.0% on RoboMME and 83.8% across four real-world tasks spanning two robot embodiments. These results represent substantial improvements over the previous state of the art in all three evaluation settings.

As multimodal large language models (MLLMs) reason for longer, attention to the initial visual input diminishes, weakening visual grounding. Visual memory reintroduces visual evidence during reasoning. We conduct a controlled analysis of visual memory along three axes: curation, organization, and access. We find that local evidence benefits from global context, compact latent representations balance accuracy and visual-context cost, and the utility of memory access depends on the reasoning state. Guided by these findings, we propose ReMAP (Reasoning-Time Memory-Augmented Perception), which couples two complementary latent memories: a static, question-conditioned Global memory that preserves scene and cross-image context, and a dynamic Local memory that uses this context as an anchor while selecting and re-encoding region-level evidence according to the current reasoning state. Both memories return compact latent tokens inserted into the reasoning sequence, and a reinforcement-learning access policy trained with branched rollouts decides when to continue reasoning or invoke Global or Local memory. On ten benchmark families, ReMAP outperforms prior visual-memory methods on all four multi-image benchmarks, exceeding the strongest prior results on MuirBench and MIMIC by 8.38 and 14.84 percentage points. Across four backbone families, enabling memory access improves over the same trained model with memory disabled, and on shared V*Bench, CV-Bench-2D, and MuirBench questions ReMAP reduces the visual tokens entering the reasoning sequence by 51.0-76.8% relative to the native-resolution backbone. Further analyses show that Global and Local memory form distinct yet complementary latent representations. Together, these components restore the perceptual cycle by letting the reasoning state trigger targeted visual retrieval, with the retrieved evidence guiding subsequent reasoning.

Aerial vision-and-language navigation (VLN) enables unmanned aerial vehicles to execute long-horizon natural-language instructions from visual observations in complex three-dimensional environments. However, recent aerial VLN models often rely on large-scale vision-language backbones and dense visual histories, imposing substantial computation and memory costs that hinder onboard deployment. We propose LightVLN, a lightweight history-aware aerial VLN framework that combines a compact 0.5B language backbone with compact representations of both historical and current observations. LightVLN compresses each historical frame into a single token using visual features already computed by the policy. It further introduces history- and instruction-conditioned local aggregation to reduce the current observation from 256 to 32 visual tokens while preserving navigation-relevant spatial information. With up to 16 historical frames, the policy uses at most 48 observation-derived tokens. On the public OpenFly dataset, LightVLN achieves 50.93% Test-Seen and 36.14% Test-Unseen success rates (SR), outperforming the evaluated 7B language-backbone baselines on most reported metrics. It also achieves 25.83% SR on AerialVLN-S Val-Seen. In a reconstructed unseen campus, we deploy LightVLN on a DJI M350 RTK with an external Jetson Orin NX 16 GB for closed-loop onboard-compute real-to-sim hardware-in-the-loop (HIL) evaluation, achieving 14.61 Hz model inference and 11.13 Hz end-to-end decision updates. These results demonstrate the effectiveness and efficiency of LightVLN for aerial navigation.

We show that multimodal models possess strong reasoning abilities and that an appropriate harness can unlock their potential to solve tasks across diverse interactive environments. We introduce VISTA, a visual harness that gives a general-purpose multimodal model long-horizon vision. VISTA allows the model to directly perceive the environment through visual observations and maintains a lossless visual memory that preserves past observations in their original form. The model can actively retrieve these observations and reorganize its visual input as it reasons. On ARC-AGI-3, VISTA improves Claude Opus 5.0's Relative Human Action Efficiency score from 40.68 to a perfect 100.00, with the model completing all 25 public games using 57.4% fewer actions than first-time human participants. VISTA's simple design also allows it to extend naturally to diverse visual environments with minimal adaptation. Across three additional benchmarks covering a diverse range of visual games and puzzles, it substantially outperforms baselines using the same underlying model with minimal harnesses. Our results highlight VISTA's potential as a general-purpose visual harness for advancing multimodal agents in complex visual environments.

Streaming video LLMs must retain evidence before its relevance to future tasks is known and respond when sufficient evidence becomes available. The challenge is to form reusable factual memory without compromising real-time perception. We introduce OneStreamer, which jointly learns query-independent evidence recording and task response through a shared proactive generation process. Its Proactive Hierarchical Caption Memory (PHCM) produces time-grounded local-detail captions and summaries of completed events. Streaming caption targets supervise the interpretation of observed video prefixes during training. At inference, model-generated records complement a recent visual window, providing reusable factual context without revisiting historical visual features. Proactive State Transition Learning (PSTL) reduces the dominance of repeated waiting states by preserving supervision at all output anchors and selecting representative state-change and state-persistence tokens. We further develop a streaming data synthesis pipeline that aligns output content and timing with available evidence. Combining the resulting streaming captions and QA with cleaned open-source data yields OneStreamer-1M, a broad-coverage streaming video interaction dataset with over one million records spanning diverse tasks. Our 4B model achieves the best results among the compared methods across all eight evaluated streaming video understanding benchmarks. Ablations show that retaining generated captions improves historical QA without degrading real-time perception. PSTL also outperforms dense state supervision while supervising only 27.5% of annotated state tokens. Together, these results support proactive generation as a shared learning interface connecting perception, memory formation, and timely response in streaming video interaction.

Streaming vision-language models must process continuously growing video streams under a bounded compute budget, creating a persistent tension between real-time perception and long-term memory. Retrieving historical information provides a natural remedy, yet historical recall is not uniformly beneficial: unnecessary history may introduce irrelevant context into current reasoning and interfere with native real-time perception. Effective streaming memory should therefore address not only what to remember, but also when and how to access it. To this end, we introduce FlashBack, a training-free framework for selective, multi-level memory in streaming vision-language models. Before retrieving history, FlashBack draws on the semantic understanding of the frozen streaming VLM to infer whether a query calls for historical evidence. This assessment determines whether inference remains on the Native trajectory or invokes an isolated Recall trajectory. The Recall trajectory combines recent context with retrieved long-term memory through a query-local Side-KV pathway, preserving local temporal continuity without modifying the persistent Native state. We instantiate FlashBack on StreamingVLM and Mage-VL-4B and evaluate it on OVO-Bench and StreamingBench. The results show improvements on several long-horizon and memory-dependent tasks while largely preserving real-time perception, with performance competitive with strong training-based streaming methods despite requiring no additional training. Our code will be announced later.

Vision-language-action (VLA) models struggle on history-dependent manipulation tasks, where the current observation alone does not determine the action, and the policy needs a memory of the history. Existing memory methods decide what to remember by design, for example, keeping frames with large pixel changes, and show inconsistent gains across tasks. We view what to remember as an optimisation problem. From the POMDP formulation of imitation learning, we show that the optimal memory maximises the conditional mutual information $I(a_t; m_t \mid o_t)$ between the action and the memory given the current observation. Intuitively, this means preserving the action-relevant information in the history that is not already contained in the current observation. Based on our analysis, we propose Divide-and-Remember (D&R), a recursive memory method that learns a memory function $m_t = M(h_t)$ and scales to long contexts while staying compute-light. It involves two strategies: (1) the selection over the full history is divided recursively into subproblems of top-$K$ selection over $2K$ tokens, so that fixed-size, lightweight selectors learned end-to-end support an unbounded history; (2) all recursion blocks share one selector, which captures the selection rule common to every block and keeps the method efficient. On RoboMME, a benchmark of 16 long-horizon manipulation tasks that require remembering when, where, what, and how to act, D&R achieves a state-of-the-art average success rate with consistent gains across all four suites under a budget of only 64 tokens; real-robot experiments show the same gain. Code, checkpoints and more results are at https://dnr-memory.github.io/

Vision-language-action policies often see only one or a few recent frames, which makes it difficult to evaluate how they use information that disappears during a task. We introduce MIKASA-Robo-VLA, a benchmark of 90 language-conditioned manipulation tasks. All but 10 hide the cue an action depends on. Those 10 are reactive controls. MIKASA-Robo, the suite it rebuilds, has 32 tasks and uses language only in a representative VLA subset. Here every task provides an instruction, while memory-dependent tasks hide a task-relevant cue and reactive controls keep it available. For 70 tasks, environment phase timings specify an information gap, and for 28 of them the gap exceeds the 16-frame window of the widest fixed-context VLA we survey. The gap counts only the interval the cue is provably absent, not the full duration a policy must retain it, so every memory-dependent task still requires memory by construction, including the ones whose measured gap is short. We release 22,500 oracle trajectories across 10 memory types in RLDS and LeRobotDataset v3. A reference $π_{0.5}$ baseline with current images and proprioception, but no observation history or explicit memory module, is fine-tuned on 14 tasks and achieves 0.211 $\pm$ 0.044 mean task success. Its lower success on the evaluated Long-split tasks is confounded by open-loop chunking and the memory types represented in that subset. Project page: https://mikasarobo.github.io/

While recent Multimodal Large Language Models (MLLMs) have shown promise in affective computing, their reasoning capabilities are largely confined to short video clips with limited interactions. However, real-world emotions are not merely isolated instantaneous reactions but dynamic and cumulative processes deeply shaped by past experiences and ongoing events. To bridge this gap, we introduce LongEmoBench, a benchmark dedicated to emotion understanding and reasoning in long videos. It assesses progressive capabilities scaling from continuous scene interactions to complex episodic developments. Furthermore, we propose LongEmo, a novel memory-augmented agentic framework designed to tackle the immense challenges of long-range affective reasoning. LongEmo processes continuous video streams to construct an Event Memory Graph, explicitly modeling long-range dependencies and capturing emotional dynamics across discrete events. Given a question, the agent retrieves a query-relevant event stream from the graph, iteratively integrating multimodal memories and relational dependencies to deduce the final answer. Extensive evaluations of 17 representative methods reveal that they struggle significantly with emotion understanding and reasoning in long videos. In contrast, LongEmo achieves state-of-the-art performance, demonstrating the efficacy of its event-centric memory architecture.

Vision-Language-Action (VLA) models have shown strong promise for general-purpose robotic manipulation, yet adapting them to new tasks and domains remains inefficient: existing methods often rely on parameter tuning, incurring substantial costs and risking catastrophic forgetting of previously learned tasks. To address this, we propose \textbf{Optimus-R}, a memory-centric VLA framework that formulates robotic adaptation as explicit query-skill memory tuning. Optimus-R introduces: (i) An \textbf{Inline Memory Interface for skill extraction}. It inserts learnable memory tokens into the VLA prefix stream, allowing the backbone to derive control-aware query and skill representations within the native action-conditioning pathway. (ii) A \textbf{Query-Skill Memory Bank for skill learning}. It externalizes skills into query prototypes for deciding \emph{what} to retrieve and skill values for specifying \emph{how} to act, supporting skill reuse and expansion with limited parameter updates. (iii) A lightweight \textbf{Bridge-and-Adapt mechanism for skill updating}. It aligns target-domain queries and skills with the existing memory space through a lightweight adapter and residual memory updates. Experiments on in-domain adaptation, cross-domain adaptation, and lifelong learning show that Optimus-R enables data-efficient skill learning while mitigating catastrophic forgetting.

Streaming video understanding requires models to process unbounded visual streams while preserving rich visual semantics across vast temporal horizons, posing a fundamental challenge for memory modeling. Existing approaches primarily focus on increasing memory capacity, either by compressing historical information into fixed-size representations or by extending storage beyond GPU memory. However, these methods largely rely on global or coarse-grained representations, inevitably losing fine-grained visual information. In this work, we argue that streaming video memory should explicitly encode structured and semantically meaningful representations, particularly at the entity level. To this end, we propose MEMO, a novel framework that models streaming video through multi-level, entity-aware structured memory. MEMO performs multi-level perception to jointly capture global semantics, entity dynamics, and spatial structures, partitioning streaming video into semantically coherent chunks. Each chunk is organized into a structured memory, where lightweight global and entity-level representations serve as retrieval indices, while the corresponding high-resolution visual content is retained separately for on-demand access. At inference time, MEMO performs query-specific retrieval over the structured memory and selectively recalls relevant visual evidence for downstream reasoning. Notably, MEMO is training-free and plug-and-play with existing multimodal large language models. Extensive experiments on StreamingBench and OVO-Bench demonstrate that MEMO consistently improves multiple base models and achieves state-of-the-art performance.

Vision-Language-Action (VLA) foundation models have recently emerged as one of the prevailing solutions for autonomous driving, as they can utilize knowledge acquired during vision-language pretraining for accurate and interpretable driving. However, VLAs can take only a limited number of frames as visual input due to the high token cost of an image, which is problematic for memory-dependent tasks such as determining the arrival order at all-way stops and long-horizon driving scene understanding. Existing solutions use latent vector memories accessed through cross-attention, which are neither interpretable nor portable. In this paper, we propose AD-Memo, a general-purpose VLA driving agent with language-based memory. The agent outputs memory as an extension of its Chain-of-Thought (CoT) to record surrounding objects critical to driving; this memory becomes part of the agent's future input. We curate memory-based datasets and train VLAs with a two-stage recipe: Supervised Fine-Tuning (SFT) and \textit{Da Capo}, a novel semi-closed-loop Reinforcement Learning (RL) algorithm which uses trajectory-level advantage for memory and step-level advantage for driving, leading to better credit assignment. Across scenarios such as all-way stops and general driving, AD-Memo improves driving quality, enables better question answering on driving scenes, and produces plug-and-play memory for other models.

Long-form video understanding requires multimodal agents to iteratively gather evidence over many reasoning steps. However, most existing agentic methods suffer from semantic thrashing: as append-only working memory grows, attention to key evidence collapses, and the agent loses access to what it has already found. First, we provide a structural argument showing that append-only memory can incorporate newly observed target evidence, but cannot remove accumulated noise or prevent ordered context growth without a rewrite operator. Second, motivated by this analysis, we propose VideoLoop, a multimodal agent with two coupled loops. The outer loop reasons over the video and the inner loop, after each step, retrieves artifacts from an unbounded filesystem of past observations and intermediate analysis, and rewrites a bounded working memory. Extensive experiments demonstrate the effectiveness of VideoLoop, which improves four popular LVLM backbones in a plug-and-play manner, with an average gain of 4.2% points over baseline on VideoMME (long). Further analysis of working memory suggests that VideoLoop mitigates semantic thrashing: on the hardest quarter of VideoMME (long) questions, a blind judge that reads only the agent's context answers 81.1% correctly, versus 60.9% for the append-only agent. With Gemini 3.1 Pro, VideoLoop reaches 88.3% on VideoMME (long), 88.8% on VideoMMMU, and 80.9% on LongVideoBench (long).

Vision-language-action (VLA) models have driven rapid progress in robotic manipulation, demonstrating strong fine-grained control and promising performance on long-horizon tasks. However, many existing VLAs lack explicit access to interaction history, making them vulnerable to perceptual aliasing: similar current observations and robot states at different task stages may induce action ambiguity and lower success rate. Existing methods incorporate temporal or progress cues through feature conditioning, action-prior modification, or sampling guidance. However, methods that jointly fine-tune memory modules and the base VLA incur additional policy-training costs, motivating the separation of trainable history-conditioned steering from frozen base-policy refinement. We propose ActMem-VLA, a dual-expert handover architecture that augments a frozen, fine-tuned VLA with a memory plugin comprising a Mamba-based memory module and a lightweight PreAction Expert (PAE). Specifically, Mamba encodes executed-action history into memory that conditions PAE alongside current context. With these inputs, PAE steers task progression during early, high-noise denoising, then passes the partially denoised action to the frozen Action Expert (AE) to refine action details during the remaining low-noise steps. The fine-tuned base VLA remains frozen throughout training, while only the Mamba module and PAE are jointly optimized. On LIBERO-Mem, ActMem-VLA achieves 80.8% average success across all ten tasks, compared with 65.2% for $π_{0.5}$ and 49.5% for MemoryVLA, while introducing only 3.45% additional parameters. Across four real-world tasks, it improves the average success rate over $π_{0.5}$ by 28.8%.

Query-agnostic streaming video understanding requires vision-language models to continuously compress an indefinitely growing visual stream into a bounded memory before future queries are known. The performance depends critically on the memory mechanism--what observations to preserve, how to represent and consolidate them, and what information to retrieve when a query eventually arrives. Rather than designing a single memory architecture by hand, we formulate memory design as a search problem over executable memory programs. We introduce a lightweight domain-specific language that expresses memory mechanisms through structured primitives for representation, admission, retention, consolidation, budgeting, and retrieval, while enforcing causal and bounded-memory constraints. Although structured, the derived program space remains large and contains heterogeneous, conditionally dependent design choices whose effects can only be assessed via downstream execution. We therefore propose MemEvo, an LLM-driven auto-research framework that uses pretrained LLM as a semantics-aware proposal model to iteratively generate and refine candidate memory programs based on accumulated experimental feedback. At runtime, a deterministic evaluation pipeline validates and evaluates each candidate, while the underlying vision-language model remains frozen throughout discovery. We finally produce a training-free, bounded-memory mechanism. Extensive experiments on StreamingBench and OVO-Bench demonstrate strong streaming video understanding performance together with substantial context and inference efficiency.

Long video understanding relies on video memory to overcome the context limits of multimodal large language models. Existing methods follow a build-then-reasoning pipeline: memory is built offline for the entire video, then reasoned over as a static source. In practice a long video is shared by several questions, and this pipeline is costly at both ends: with few questions, building memory for the whole video costs far more than answering them; with many questions, the memory is never updated, so what is learned while answering questions is lost to the next question. To alleviate these, we introduce Sprout, an agentic framework that builds memory while reasoning: a temporal tree that sprouts detailed nodes as questions are answered. The agent watches the video segment by segment at a low frame rate, stopping when the current question can be answered, remembers each segment as a coarse node of the tree, and revisits key intervals at a higher frame rate to refine the tree with the recovered details. Once a segment is recorded as text, its video input is removed from the context history, while the original video remains reachable through the video tools. The memory tree and prior question--answer records persist across questions, so the memory is online and dynamic: built from the first question onward and updated by every question thereafter. We find that replacing accumulated video inputs with textual memory substantially reduces context usage while maintaining accuracy, with slight improvements in some settings. Across benchmarks on three models, Sprout achieves competitive or improved accuracy relative to representative offline memory methods, with no upfront construction stage and lower context cost per question.

Long-horizon manipulation requires robots to remember cues that are no longer in view while responding to moving objects. Yet vision-language-action (VLA) policies often rely on the latest observation, and refreshing their visual context typically requires another costly vision-language model (VLM) pass. We present D$^2$-VLA, which combines dual memory and dual-frequency control at the KV-cache interface of a pretrained VLA. D$^2$-VLA uses block-wise causal KV caching to encode observations incrementally and, guided by distinct temporal attention patterns, constructs separate historical KV read views for the VLM and action expert. Between periodic VLM updates, a gated adapter incorporates fresh visual features into the latest history-conditioned KV block, while a short fast-memory queue supports action replanning. We introduce DOMINO-Long, a ten-task benchmark requiring robots to use earlier visual cues when manipulating moving objects. D$^2$-VLA achieves complete-task success rates of 29.3% on DOMINO, compared with 9.6% for $π_{0.5}$ and 17.2% for PUMA, and 60.0% on DOMINO-Long, compared with 35.4% and 20.6%, respectively. It improves success rates on eight real-robot tasks and reaches 97.5% on LIBERO-Long and 74.3% on RoboTwin 2.0.

Memory is essential for long-horizon, partially observed robotic manipulation: a robot must remember which object was placed in a drawer, whose cup it moved, or how many action cycles have elapsed. Recent vision-language-action (VLA) models embed memory directly inside the policy, but benchmarks show no single in-policy mechanism covers all spatio-temporal dimensions, trailing oracle methods by a wide margin. We argue that memory type dictates where memory should reside: short-term perceptual memory (repetition, timing, retracing) belongs inside the policy, while long-term object memory (persistent spatial state, containment, event history) belongs outside as an explicit, readable record. We present Ledger, a harness that realizes this split over a single fine-tuned $π_{0.5}$ policy by pairing an in-policy frame-sampling memory with an external spatio-temporal object memory, the ledger, built from a SAM3 tracker and a VLM captioner of the demonstration and read by an LLM planner that decides at step boundaries. On RoboMME, Ledger reaches the highest four-suite average among the evaluated methods, 64.3% (vs. 45.9% for the strongest prior method under identical evaluation), leading object reference (60.7% vs. 40.3%) and object permanence (86.7% vs. 56.2%) using a single set of weights. Choosing the memory source at runtime, from the instruction and the record, removes the need for a task-level router.

Active perception and manipulation are crucial for robots to interact with complex scenes. Existing benchmarks struggle to evaluate how robots effectively acquire and maintain information in memory in an active manner. To this end, we introduce ActiveArena-Sim, an active-perception simulator with controllable viewpoints and large-scale workspaces as the foundation. Built on this, we propose ActiveArena-Bench, which comprises 35 tasks across 5 fine-grained categories, covering visual exploration and interactive information acquisition. Each task is difficult to solve from passive observations alone, requiring multi-round evidence acquisition and memory-based reasoning. The benchmark provides rich memory annotations, standardized training data, and ID/OOD protocols featuring disjoint scenes, unseen distractor configurations, and novel backgrounds. Moreover, we present ActiveArena-VLA, a modular suite of 13 vision-language-action configurations for controlled studies of memory writing, memory capacity, proprioceptive state, subtask supervision, and high-level planning in active perception. Benchmark results reveal a substantial ID-OOD gap: uniform memory sampling, increased memory capacity under reliable write policies, proprioceptive inputs, and subtask supervision improve OOD generalization, while planner-guided memory management and decision-making achieve performance close to the best-performing configuration using only sparse memory. ActiveArena thus provides a unified testbed to develop and diagnose models for active perception and manipulation.

Long-video understanding must capture transient visual evidence under strict token budgets, yet existing methods compress frames, append memory tokens, or alter internal key-value (KV) caches. We introduce Prefix-Steered Recurrent Memory (PREM), a memory-token-free framework for frozen vision-language models (VLMs). PREM separates video ingestion from query answering: a recurrent writer distills visual streams into a compact 256 KiB multi-slot associative state, while a question-conditioned readout adds memory-derived key/value (K/V) steering modulations to existing non-visual prompt prefixes during prefill. This enables write-once, query-many inference without extra prompt tokens or decoding recurrence. Across six long-video benchmarks in offline and streaming end-of-stream settings, PREM consistently outperforms frozen baselines at every evaluated visual budget. Under a constrained budget of 16 frames, PREM improves macro-average accuracy by 3.06% on Qwen2.5-VL-3B, with gains of 11.0% on action antonym identification and 9.9% on localized needle retrieval. These gains require tuning 0.24% of backbone parameters at 0.03 GiB of peak GPU memory overhead.

Reactive vision-language-action (VLA) policies suffer from task-state aliasing in long-horizon manipulation, where identical multimodal inputs call for distinct, context-dependent actions. Given that pretrained VLAs already possess rich control primitives to express diverse behaviors, we hypothesize that the execution bottleneck lies not in policy capacity, but in input ambiguity. In this paper, we propose TaskAnchor, a lightweight adapter that grounds task state by injecting execution context into the VLA's native input space. During post-training, TaskAnchor learns to represent the semantic execution stage as a milestone-supervised coordinate prepended to the language instruction, while incorporating fine-grained historical evidence via a residual update to the current visual tokens. This formulation avoids generating complex subtask instructions and leaves the backbone architecture unchanged. Across long-horizon benchmarks, TaskAnchor delivers substantial gains, achieving approximately 6 times the average success rate of the pi0.5 and X-VLA baselines on RMBench and more than doubling the task success rate of pi0.5 on RoboMemArena. Real-robot experiments further validate reliable multi-stage execution, with the same policy adapting its subsequent behaviors using earlier human interactions as in-context cues. Our project website is available at https://taskanchor.netlify.app/.

Cache-based test-time adaptation improves CLIP predictions by storing and retrieving examples from the target stream while keeping the model frozen. However, existing methods largely treat the feature space used for image-image retrieval as fixed, leaving open how much adaptation depends on the retrieval space itself. We study this question by fixing the memory and changing only the retrieval encoder, finding that the same memory can yield very different gains: across sixteen retrieval spaces, ImageNet-A cache gain ranges from at most +0.44 points for CLIP and MAE to +19.7 +/- 0.4 for DINOv2-L, while label-free retrieval-space selection retains 98% of oracle gain on ImageNet-V2. These results show that memory quality depends not only on which examples are stored, but also on how they are retrieved. Motivated by this finding, we propose MARC (Memory Augmented Retrieval for CLIP), a training-free system that uses frozen CLIP for prediction and DINOv2-B for retrieval with a single fusion weight. A single-view cache repairs 1074 +/- 21 baseline errors, compared with 878 +/- 4 for a 64-view ensemble, at roughly one seventh of the cost. Across four ImageNet distribution shifts, MARC reaches a 67.91% OOD average and, at matched DINOv2-B scale and eight views, achieves 64.17 +/- 0.31% versus 62.75 +/- 0.15% for a graph-based cache system while running 2.6 times faster. Overall, our results establish retrieval space as a first-order design choice for robust cache-based adaptation in remote sensing, scientific imaging, and changing visual environments.

State-of-the-art vision-language-action models (VLA) for autonomous driving face critical limitations: excessive parameter counts, inefficient high-resolution image processing, and lack of temporal memory. We introduce Fast and EffectIVE VLA (FIVE-VLA) to address these through two key contributions. First, we employ an efficient vision encoder that processes high-resolution ($448 \times 896$) images while generating only 98 tokens, over $5\times$ fewer than existing approaches, and bypass text generation entirely for single-pass trajectory prediction. Second, we propose Recurrent Action Memory (RAM), a lightweight module that conditions action prediction on previous action tokens, providing temporal context critical for manoeuvres such as overtaking and emergency braking. With only 641M parameters, FIVE-VLA completes $\sim$10% more routes without traffic rule infractions than the previous state-of-the-art VLA on the challenging Bench2Drive closed-loop driving benchmark. Non-reactive open-loop simulation on the large-scale real-world NVIDIA Physical AI AV dataset shows 10.2% and 7.7% lower collision-violation rates than SimLingo in single- and four-view settings, respectively. Additionally, FIVE-VLA runs at $\sim$30 fps on an A100 and $\sim$4 fps on a T4 GPU (proxy to an edge device), representing an 8-30$\times$ speedup over previous methods.

Vision-language model (VLM) agents combine specialized perception, tools, and reasoning to address complex visual tasks. In multi-agent settings, different agents inspect different image regions, video frames, or visual representations, so collaboration extends beyond distributed reasoning to distributed perception. This makes shared visual context a central problem in VLM agent collaboration. In this paper, we frame memory hierarchy, cross-agent sharing, and consistency mechanisms around the need to reconcile interpretations and update dependent reasoning. Effective collaboration requires agents to build on contributions from other agents, recover missing visual context, and reconcile differing interpretations as new evidence emerges. Shared visual memory preserves not only images or textual summaries but also the dependencies among observations, agent interpretations, and subsequent reasoning. Together, these design considerations shape how information flows and evolves across VLM agents. The proposed framework provides a foundation for building reliable and resource-efficient agent teams.

Wearable assistants require episodic memory over egocentric video, yet current vision-language models face bounded frame budgets, growing visual-token costs, and long-context retrieval failures. Under these practical constraints, we study whether textual captions can serve as reusable episodic memory. We define the Episodic Memory Video Caption QA task and introduce CapMem, a human-annotated benchmark with 75 videos totaling 33.7 hours, and 1,000 multiple-choice questions across 16 scenarios. On long videos (>20 min), full-coverage CaptionQA with 30s and 60s caption windows outperforms direct VideoQA for 10/12 and 8/12 models, respectively. On the same video subset, a matched-frame control across six Qwen models retains mean accuracy gains of 3.22 and 2.55 points, respectively. Our caption-guided retrieve-and-verify harness further improves accuracy by up to 5.3 points. These results support the effectiveness of caption memory for episodic reasoning over long egocentric video.

Vision-Language Models (VLMs) enable autonomous GUI navigation, but agents still struggle to process and learn from dense, continuous visual histories. This bottleneck hinders both immediate error correction within a single episode (intra-trial) and experience distillation across multiple attempts (cross-trial). We trace these challenges to an empirical informational asymmetry in GUI navigation: while expected transitions can often be compressed into lightweight textual summaries, unexpected outcomes benefit from preserved screenshots as causal evidence for accurate diagnosis. Building on this insight, we propose AnchorGUI, a unified framework driven by the Cognitive State Anchor (CSA). The CSA acts as a per-step primitive that actively compares expected and observed transitions, converting passive multimodal trajectories into explicit prediction-error signals. These signals orchestrate a dual-scale learning mechanism via an asymmetric memory. For intra-trial correction, a sliding window selectively retains visual evidence for detected mismatches, providing immediate, visually-grounded feedback. For cross-trial distillation, this asymmetric memory focuses the computationally expensive credit assignment search space on likely failure steps. Experiments across four benchmarks validate the effectiveness of our approach. On AndroidWorld, AnchorGUI achieves a 57.3% success rate with a $2.4\times$ token reduction per step. Furthermore, cross-trial distillation reaches 69.2% success (+11.9% gain), significantly outperforming standard reflection methods while maintaining sub-linear context scaling.