RoboMemory in Embodied Robotics

• RoboMemory is a comprehensive robotics paradigm that treats memory as a fundamental, multi-modal component essential for grounding sensorimotor and semantic experiences.
• It employs explicit long-term storage, dynamic retrieval, and structured benchmarks to evaluate memory robustness under partial observability and interference.
• Empirical studies demonstrate that integrating diverse memory taxonomies enhances control efficiency and adaptability in complex, long-horizon robotic tasks.

In the robotics literature, RoboMemory denotes both specific memory-augmented systems and a broader research agenda that treats memory as a first-class component of embodied intelligence rather than as an auxiliary cache. The term encompasses long-term queryable stores for sensor experience, distributed episodic memory in cognitive architectures, dynamic spatial-semantic world models, recurrent and belief-state memory for control under partial observability, and benchmark suites that make history dependence measurable. Representative instantiations include the brain-inspired "RoboMemory" framework with Spatial, Temporal, Episodic, and Semantic memory (Lei et al., 2 Aug 2025), the long-term query-answering system "RoboMem" (Idrees et al., 2020), and standardized evaluation suites such as RoboMME and RoboMemArena (Dai et al., 4 Mar 2026, Lei et al., 11 May 2026).

1. Conceptual scope and memory taxonomies

A recurrent theme across the literature is that robot memory is not a single mechanism. In the ArmarX cognitive architecture, memory is defined as a central active component that mediates between semantic and sensorimotor representations, orchestrates data streams and events between processes, and supports abstraction, symbolic plan parametrization, and prediction. The same work identifies five conceptual requirements: memory should be active, multi-modal, associative, introspective, and inherently episodic (Peller-Konrad et al., 2022). In the 2025 RoboMemory framework, this broad view is operationalized as four explicit memory types—Spatial, Temporal, Episodic, and Semantic—within the loop “Perception -- Memory -- Retrieval -- Planning -- Execution,” with an Information Preprocessor, a Comprehensive Embodied Memory System, a Closed-Loop Planning Module, and a Low-level Executor (Lei et al., 2 Aug 2025).

Benchmark papers formalize related but distinct taxonomies. RoboMME organizes history dependence into temporal, spatial, object, and procedural memory, corresponding to “when,” “where,” “what,” and “how” questions in long-horizon manipulation (Dai et al., 4 Mar 2026). MIKASA instead classifies memory-intensive RL tasks into Object Memory, Spatial Memory, Sequential Memory, and Memory Capacity, reflecting partial observability, delayed dependence, and multi-item retention in tabletop manipulation (Cherepanov et al., 14 Feb 2025). RoboMME-Interference further refines the problem by distinguishing between “can the policy use memory at all?” and “can the policy still use memory when unrelated sessions intervene?”, thereby making interference and cross-session robustness explicit evaluation targets (Rathi, 21 Jun 2026).

These formulations converge on a common point: memory in robotics is not reducible to extending an observation window. The task may require counting, referential grounding, occlusion reasoning, trajectory imitation, recall across sessions, or compression of multimodal histories into action-sufficient state. This suggests that “RoboMemory” is best understood as a family of representational and algorithmic choices rather than a single canonical module.

2. Long-term storage, episodic structure, and queryable memory

One major branch of RoboMemory research treats memory as an explicit, persistent, queryable store. RoboMem was proposed as a long-term memory system for robots assisting elderly people over days, months, or years of data. Its pipeline performs real-time preprocessing on video and SLAM pose information, stores compact structured memory in a database, updates that memory through successive refinement, and falls back to reprocessing only a small set of relevant raw frames when the database alone is insufficient (Idrees et al., 2020). The system translates natural-language queries into an intermediate representation $Query'$, converts that into database commands, and retrieves the highest-probability attributes from the database. In the reported prototype, RoboMem stores 3.5 MB in the database versus 535.8 MB for the actual video frames plus metadata, with an average DB query time of 0.0002 s and Detectron inference time of 0.143 s per image on average (Idrees et al., 2020).

ArmarX generalizes this idea into a distributed cognitive memory system. Its architecture is organized as a federation of memory servers registered in a Memory Name System (MNS), with the full memory defined as the union of all distributed servers (Peller-Konrad et al., 2022). Internally, memory is structured hierarchically as Memory $\rightarrow$ Core Segment $\rightarrow$ Provider Segment $\rightarrow$ Entity $\rightarrow$ Snapshot $\rightarrow$ Instance, making every stored object part of an episodic timeline rather than a flat fact table. The framework introduces the ArmarX Interpretable Data Format (IDF) for introspectable and language-agnostic storage of symbolic and sub-symbolic content, and supports both event-driven notifications and query-based access. Its long-term memory pipeline filters, compresses, and later reconstructs stored data; in the reported ARMAR-III pick-and-place recordings, an unfiltered data rate of 35.665 MB/s dropped to 0.213 MB/s after online compression and to 0.00383 MB/s after offline auto-encoder compression (Peller-Konrad et al., 2022).

These explicit-memory systems differ from latent recurrent memory in two respects. First, the stored content remains inspectable: frame numbers, labels, poses, timestamps, object locations, or typed IDF objects can be queried directly. Second, consolidation and refinement are algorithmic primitives rather than side effects of gradient descent. A plausible implication is that explicit RoboMemory architectures are especially appropriate when later interrogation, explanation, or offline analysis is as important as immediate control.

3. Benchmarks and the measurement of memory robustness

The benchmark literature has shifted robot memory from an informal capability claim to a measurable variable. RoboMME introduced a standardized manipulation benchmark with 16 tasks, 1,600 demonstrations, and 770k timesteps, designed under a taxonomy of temporal, spatial, object, and procedural memory (Dai et al., 4 Mar 2026). It also released a controlled family of 14 memory-augmented VLA variants on the $\pi_{0.5}$ backbone, spanning symbolic, perceptual, and recurrent memory with different integration mechanisms. Under a fixed memory budget $B=512$, the strongest non-oracle model is FrameSamp + Modul at 44.51% average success, while GroundSG + Oracle reaches 84.08%, showing that high-quality symbolic abstractions can be extremely effective when the subgoals are accurate (Dai et al., 4 Mar 2026).

RoboMME-Interference extends this line by making long-context interference the central variable. For each query episode, it constructs a session history consisting of the relevant prior demonstration followed by a controlled number of unrelated distractor sessions, with $k \in \{0,1,3,7\}$, and measures success when that ordered history is provided to the VLA as memory (Rathi, 21 Jun 2026). Each distractor unit is 32 stored frames, specifically “the first 256 raw frames of the distractor episode subsampled at stride 8,” and distractors are drawn from different task families to isolate interference rather than contradiction. Across 9 systems, 9 task families, and 50 test episodes per family, the paper reports that the strongest overall system, FrameSamp-Modul, rises from a no-history floor of 18.2% to 45.3% at $k=0$, then falls by about 26 points by $\rightarrow$0, ending near its no-history performance. TokenDrop-Modul shows the same qualitative pattern, while Recurrent-TTT-Expert and Recurrent-TTT-Context remain near the no-history floor across all interference levels (Rathi, 21 Jun 2026).

RoboMemArena enlarges the scope further. It contains 26 simulated tasks, 151 distinct subtasks, 104/151 memory-dependent subtasks, an average trajectory length: 1,076 steps per task, 2,600 successful demonstrations, and 15,100 keyframe-aligned short segments for hierarchical supervision (Lei et al., 11 May 2026). Its paired model, PrediMem, is a dual-system VLA with a recent buffer, a keyframe buffer, and a predictive coding head. On RoboMemArena, PrediMem reports 38.5 TSR / 55.2 CSR, compared with 21.5 TSR / 38.7 CSR for $\rightarrow$1; on the physical benchmark it reaches 52% avg, compared with 20% avg for $\rightarrow$2 and 40% avg for MemER (Lei et al., 11 May 2026).

MIKASA-Robo complements these benchmarks from the RL side. Built on ManiSkill3, it provides 32 memory-intensive robotic tabletop manipulation tasks grouped into 12 categories and supports State, RGB, RGB + joints, Oracle, and Prompt observation modes (Cherepanov et al., 14 Feb 2025). Its validation result is diagnostically important: in state mode with dense reward, PPO-MLP achieves 100% success on all tasks, while in RGB + joints mode both PPO-MLP and PPO-LSTM deteriorate sharply as task complexity increases, especially under sparse rewards (Cherepanov et al., 14 Feb 2025).

Taken together, these benchmarks undermine two common simplifications. First, memory is not equivalent to a longer recent context window. Second, success on short-horizon or single-episode tasks does not imply robustness to long histories, irrelevant prior sessions, or memory-dependent task composition.

4. Partial observability, selective recall, and memory-efficiency constraints

A second major strand of RoboMemory research studies memory as a control primitive under partial observability and hardware constraints. In active memory reduction (AMR), the goal is not merely to store history but to learn policies that actively seek to reduce their own memory requirements. The method jointly learns the memory update $\rightarrow$3 and the policy $\rightarrow$4, relaxes memory minimization with a group LASSO penalty on the memory-layer weights, and optimizes the resulting objective by policy gradient (Booker et al., 2020). Empirically, AMR-PG consistently recovers the memory-optimal two-state policy in the discrete navigation example, reduces memory dimension from 300 to at most 4 in the continuous maze, and reduces memory from 100 dimensions to 34, a 66% reduction, in the iGibson apartment task (Booker et al., 2020).

MEMBOT addresses a different failure mode: intermittent observability, where the agent receives $\rightarrow$5 with $\rightarrow$6 and $\rightarrow$7 (Liang et al., 14 Sep 2025). Its architecture separates an observation encoder, a memory-based observer or belief encoder, and a task-specific policy, and trains them in two phases: offline multi-task pretraining with behavior cloning and reconstruction, followed by task-specific fine-tuning with SAC. On 10 MetaWorld and Robomimic manipulation tasks, the paper reports that MEMBOT maintains up to 80% of peak performance under 50% observation availability; on Drawer-Close, MEMBOT-LSTM remains around 65.8% success at $\rightarrow$8, whereas memoryless baselines are around 37% (Liang et al., 14 Sep 2025).

Selective use of history is the focus of Gated Memory Policy (GMP). GMP argues that simply extending observation history often causes distribution shift, overfitting, and quadratic self-attention cost, so it learns a binary gate $\rightarrow$9 for deciding when to recall memory and a lightweight cross-attention module for deciding what to recall (Gao et al., 21 Apr 2026). It further injects diffusion noise into historical actions during both training and inference to improve robustness. On MemMimic, GMP achieves a 30.1% average success rate improvement over long-history baselines, while on Markovian RoboMimic tasks it remains competitive. The gate is usually off even on memory tasks—about 73% off in Match Color and about 58% off in Iterative Pushing—and the long-delay Match Color setup reaches 99.0% ± 1.0% success with a 6000-frame / 6000-action memory buffer and only 0.16s inference for 8 denoising steps on an RTX 5090 (Gao et al., 21 Apr 2026).

AURA-Mem makes the resource argument explicit. It replaces a growing Transformer KV-cache with a bounded fast-weight recurrent memory $\rightarrow$0 and a learned gate that writes only when the current observation would change the next action (Chen, 1 Jun 2026). Its inference state is fixed at 4,224 bytes regardless of horizon, while a matched KV-cache grows to 25,600,000 bytes at 100k steps, about 6,061\times larger. On the synthetic noisy_long_recall benchmark, AURA-Mem matches the best $\rightarrow$1 baseline in accuracy while using 5.19-6.13 times fewer writes, and up to 9.19 times fewer writes on easier configurations. On a trained closed-loop OpenVLA-OFT 7B panel on LIBERO-Long, it matches the ungated base policy at 0.233, exceeds an always-write KV arm at 0.217, and uses 7.0 times fewer writes (Chen, 1 Jun 2026).

These results support an increasingly clear conclusion: more memory is not automatically better, and recurrent state is not automatically effective. What matters is whether the stored information is action-sufficient, whether it degrades gracefully under missing observations or distractors, and whether the write/read mechanism is compatible with the bandwidth and latency regime of embodied hardware.

5. Spatial-semantic world models and multi-memory embodied agents

Another major meaning of RoboMemory concerns persistent world models for navigation and mobile manipulation. The 2025 RoboMemory framework implements Temporal Memory as a bounded FIFO buffer of summarized interaction steps, Spatial Memory as a dynamic knowledge graph $\rightarrow$2 over objects and spatial relations, and Episodic and Semantic Memory as RAG-style long-term stores that consolidate task histories and distilled lessons (Lei et al., 2 Aug 2025). Spatial updates are localized by retrieval and $\rightarrow$3-hop expansion, after which a VLM-based resolver updates only the induced local subgraph rather than the entire graph. On EB-ALFRED, RoboMemory reports 67.0% average SR and 78.4% average GC, improving average SR by 25% over its baseline and exceeding Gemini-1.5-Pro by 3%; ablations show substantial drops without the critic (55% average SR), without spatial memory (47%), or without long-term memory (57%) (Lei et al., 2 Aug 2025).

DREAM targets the same long-horizon problem in dynamic indoor environments, but with a different representational substrate: a dynamic spatio-semantic voxel memory registered to a LiDAR-inertial-visual SLAM backend (Yan et al., 30 May 2026). Each occupied voxel stores 3D position, observation count, source image ID, semantic feature vector, and latest observation time; stale voxels are removed by depth-consistency tests, and Redundancy-Aware Memory Pruning (RMP) keeps the map bounded under pose-graph corrections. DREAM combines language-conditioned 3D retrieval, OWL-V2 open-vocabulary detection, and Qwen-VL-Max semantic verification. In real-robot experiments across four dynamic scenes, it improves long-horizon success from 40%-60% with DynaMem to 55%-70%, with an aggregated rise from 39/80 = 48.8% to 50/80 = 62.5%, while keeping a memory footprint of 0.37-0.63 GB and online memory-update time of 0.43-0.53 s (Yan et al., 30 May 2026).

Meta-Memory addresses a narrower but technically demanding task: spatial localization question answering. It builds a high-density semantic-spatial memory in which each time chunk stores a VLM-generated caption, a text embedding, a composite image formed from four evenly spaced frames, and the robot’s position (Mao et al., 25 Sep 2025). Retrieval is split into Semantic-Similarity Retrieval (SSR), Spatial-Range Retrieval (SRR), and Memory-Integration (MI), with the last stage constructing a cognitive map from landmark positions and Dijkstra shortest paths. On SpaceLocQA, which contains 6 handheld video sequences, 68 minutes, and 270 total queries, Meta-Memory reaches 67.8 on basic, 61.8 on local, and 62.2 on global questions, outperforming ReMEmbR, Embodied-RAG, and human baselines; on NaVQA it reduces mean positional error to 21.7, compared with 28.5 for ReMEmbR and 31.1 for Embodied-RAG (Mao et al., 25 Sep 2025).

Robo-Cortex extends the memory idea from storage to self-evolution. Its Dual-Grain Cognitive Memory consists of Short-term Reflective Memory (SRM) for within-episode local progress analysis and Long-term Principle Memory (LPM) for principle-level abstractions from past trajectories, while Autonomous Knowledge Induction (AKI) distills multimodal experience into a Navigation Heuristic Library (Chan et al., 18 May 2026). On IGNav, AR, and AEQA, Robo-Cortex achieves 41.26 SR and 31.66 SPL on IGNav, 22.39 SR on AR, and 29.78 Answer Score on AEQA; the adaptive version, Robo-Cortex++, improves to 45.07 SR and 35.06 SPL on IGNav, 23.88 SR on AR, and 30.59 Answer Score with 25.57 SPL on AEQA (Chan et al., 18 May 2026).

Across these systems, memory is not treated as a passive log. It is a mutable scene representation, a retrieval substrate for language grounding, a source of principles or heuristics, and a bridge between current observations and cumulative experience.

6. Collaboration memory, data infrastructure, and forensic memory

RoboMemory research also extends beyond single-agent control. In human-robot teamwork for urban search and rescue, prior collaboration patterns are represented as knowledge-graph episodic memories and selected for reuse by an RGCN encoder with a node-classification objective (Kim et al., 17 Jun 2026). The memory pool contains 209 CPs collected from earlier MATRX studies. In experiments with 20 student participants and 160 round-level observations, initializing the robot with a single automatically selected prior collaboration pattern increases rescue success from 25.7% to 41.3% and reduces average task time by 283 seconds. The strongest improvement occurs in Round 1, where success rises from 0.0% to 55.0% and remaining rocks drop from 18.2 to 5.8, although victim harm increased on average in the memory condition and performance drops in Rounds 7 and 8 (Kim et al., 17 Jun 2026). This shows that explicit episodic memory can improve early teamwork while also introducing task-mismatch and safety tradeoffs.

At the data-management layer, Robo-DM treats robot trajectories themselves as memory objects. It stores multimodal episodes in a self-contained EBML container with relative timestamps, supports lossy and lossless compression, and uses memory-mapped decoded caches with load balancing for training-time access (Chen et al., 21 May 2025). Compared to RLDS, it reports up to 70x space reduction with lossy compression and up to 3.5x space reduction with lossless compression, and compared with LeRobot it is up to 50x faster sequentially decoding. In the physical ICRT pick-and-place study, lossy Robo-DM compression reduces a 5.8 GB dataset to 77 MB, about 75.3x smaller, while the trained model achieves 15/15 successes (100%) (Chen et al., 21 May 2025). The paper is explicit, however, that Robo-DM can be slower than uncompressed HDF5 in some settings and that its RAM-heavy cache strategy can increase memory pressure.

A distinct operational meaning of memory appears in robotics forensics. For ROS-based systems, volatile memory forensics treats RAM as the primary evidence source for live incidents in which shutdown would destroy runtime state (Vilches et al., 2018). The ros_volatility project extends Volatility with a linux_rosnode plugin that starts from linux_pslist, filters processes using ROS libraries, inspects sockets, and checks whether nodes appear registered in the ROS network. In the demonstrated attack, an unauthenticated ROS Master API call unregisters a publisher node without killing its process, producing a denial of service. The paper reports the RVSS vector $\rightarrow$4 with a score of 7.6/10, classified as High severity, and shows that memory analysis can reveal talker (unregistered) together with shell-history evidence from roschaos (Vilches et al., 2018).

In this broader sense, RoboMemory is not limited to action selection or planning. It also refers to how robotic experience is stored, shared, revisited, compressed, and audited. The literature therefore spans persistent task memory, benchmarked long-context recall, bounded recurrent state, spatial-semantic world models, reusable collaboration episodes, data containers for large demonstration corpora, and volatile forensic evidence. The unifying proposition is that embodied systems deployed over long horizons require memory that is structured, updateable, and operationally meaningful—not merely larger context windows.