The DOI matches its Crossref record; 91 of the 91 references it cites were found in a registry or an index; 6 of the 8 sources its claims cite were checked for support.
KARMA: Augmenting Embodied AI Agents with Long-and-Short Term Memory Systems
Open accessAbstract
Embodied AI agents responsible for executing interconnected, long-sequence household tasks often face difficulties with in-context memory, leading to inefficiencies and errors in task execution. To address this issue, we introduce KARMA, an innovative memory system that integrates longterm and short-term memory modules, enhancing large language models (LLMs) for planning in embodied agents through memory-augmented prompting. Karma distinguishes between long-term and short-term memory, with long-term memory capturing comprehensive 3D scene graphs as representations of the environment, while short-term memory dynamically records changes in objects' positions and states. This dualmemory structure allows agents to retrieve relevant past scene experiences, thereby improving the accuracy and efficiency of task planning. Short-term memory employs strategies for effective and adaptive memory replacement, ensuring the retention of critical information while discarding less pertinent data. Compared to state-of-the-art embodied agents enhanced with memory, our memory-augmented embodied AI agent improves success rates by$1.3 \times$and$2.3 \times$in Composite Tasks and Complex Tasks within the AI2-THOR simulator, respectively, and enhances task execution efficiency by$3.4 \times$and$62.7 \times$. Furthermore, we demonstrate that KARMA's plug-and-play capability allows for seamless deployment on real-world robotic systems, such as mobile manipulation platforms. Through this plug-and-play memory system, KARMA significantly enhances the ability of embodied agents to generate coherent and contextually appropriate plans, making the execution of complex household tasks more efficient. Our code is available at https://github.com/WZX0Swarm0Robotics/KARMA/tree/master.
Overview
CoChat Summary
TL;DR
KARMA is a plug-and-play long-and-short-term memory system for embodied AI agents that augments LLM task planning with a 3D scene graph and dynamically updated object memories. In AI2-THOR/ALFRED-L, it improves success rates by 1.3× on Composite Tasks and 2.3× on Complex Tasks, and improves execution efficiency by 3.4× and 62.7×, respectively.
Key findings
- KARMA improves over memory-augmented embodied agents: success rate by 1.3× on Composite Tasks and 2.3× on Complex Tasks, and task execution efficiency by 3.4× and 62.7×, respectively. On Simple Tasks it improves success rate by 1.1× and efficiency by 2.3× over HELPER.
- Memory retrieval accuracy is higher for Composite Tasks than Complex Tasks: KARMA reports MRA 0.93 for Composite Tasks and 0.42 for Complex Tasks, described as 2.2× higher for composite tasks, with ambiguous instructions limiting semantic matching.
- For memory replacement, W-TinyLFU with a window segment size of 9 achieves the highest memory hit rate after warm-up on ALFRED-R; larger memory sizes also help, with FIFO hit rate at 25 memory units 4.6× higher than at 5 units, and W-TinyLFU 3.9× higher.
- Ablation shows short-term memory mainly affects success rate, while long-term memory mainly affects efficiency: removing short-term memory drops success rate by 1.9× and 4.2× on complex and composite tasks, respectively; removing long-term memory reduces RT by 2.7× but SR only by 1.2×.
Methods
- KARMA combines long-term memory as a non-volatile 3D scene graph of static environment structure with short-term memory as volatile units storing object type, position, state, and image, recalled via embedding similarity and fed into an LLM planner.
- Evaluation uses the AI2-THOR simulator and ALFRED-L, built from ALFRED tasks and containing 48 high-level instructions: 15 Simple Tasks, 15 Composite Tasks, and 18 Complex Tasks; ALFWorld-R is used to assess memory replacement.
- Baselines include modified LoTa-Bench, HELPER, and CAPEAM; real-world deployment uses a mobile manipulation robot with a UR3 arm and six-wheeled chassis, Google Cartographer, LangSAM, and AnyGrasp.
Limitations
- The paper states evaluations are under ideal simulation environments, without interruptions by other agents or humans, and it lacks real-world evaluation of memory behavior under larger object counts or intentional human disturbances.
- The paper states the memory design borrows terminology from human memory but lacks biological theoretical support.
- The paper states planning and memory operations are open-loop, with no feedback mechanism for evicting or updating incorrect memories; the text also makes evident that real-world deployment is demonstrated but not quantitatively evaluated in the reported results.
Model-written summary. Check it against the paper before relying on it.
Integrity signals
Registry
Cites
Statistics
Verification
Claims the paper makes
4 Valid · 3 Not valid · 0 Not checked · 61 Own results and interpretations
The cited sources support only part of this claim — read each source’s verdict below.
In the paper
“However, these approaches fail to address challenges such as hallucinations or memory inconsistencies that often arise long-sequence task planning with LLMs.”
What the sources say
- Partly supportedcites Kagaya et al., 2024 sourceExperimental — accuracy not yet measured
The Introduction says prior approaches fall short in using comprehensive past information in complex environments, supporting the general claim that they leave a planning challenge unresolved. It does not identify hallucinations or memory inconsistencies in long-sequence planning as that challenge; searches for those terms found no passage addressing them. The indexed copy provides section labels but no page numbers. (LLM, conf=0.78)
- Could not verifycites Zhang et al., 2023 sourceExperimental — accuracy not yet measured
Section 2.2.1 describes memory modules for recommender-system agents, and Appendix B.3 discusses correcting misleading preferences and the practical limits of storing long interaction histories. Neither passage addresses whether prior memory-augmented embodied agents fail to handle hallucinations or memory inconsistencies in long-sequence task planning; searches for those challenges found no supporting passage. (LLM, conf=0.76)
- Not matchedcites Sarch et al., 2023a
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
plain- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Partly supported
Every source the paper cites for this claim was checked and supports it.
In the paper
“allowing long- and short-term memory storage and recall methods to be seamlessly integrated into task planning for real robotic systems(Sun et al., 2024).”
What the sources say
- Supportedcites Sun et al., 2024 source
Section 3.4 describes a robot memory module with short- and long-term components, and the Short-term memory section explains how the planner uses recalled information to adjust actions. Section 4.5.2 says the module is integrated during planning; Section 4.1 reports evaluation with a physical robot in a real-world environment. Page numbers were not available in our copy. (LLM, conf=0.91)
plainstates a cause- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Supported
The cited sources support only part of this claim — read each source’s verdict below.
In the paper
“Using an LLM as its core planning module can effectively decompose long and complex tasks into multiple short and fixed movements (Choi et al., 2024; Sarch et al., 2024; Chen et al., 2023b; Vemprala et al., 2023; Rana et al., 2023; Brohan et al., 2022, 2023; Belkhale et al., 2024), increasing the success rate.”
What the sources say
- Partly supportedcites Choi et al., 2024 sourceExperimental — accuracy not yet measured
Section 3 (p. 3) describes an LLM planner that turns an instruction into a sequence of predefined atomic skills, and Section 5.2 (p. 6) reports some success on complex tasks requiring longer sequences. The paper does not show that decomposing tasks into these skills itself increases the success rate, nor does it characterize the skills as fixed movements. (LLM, conf=0.85)
- Partly supportedcites Sarch et al., 2024 sourceExperimental — accuracy not yet measured
Section 2.1 (page unavailable) describes an LLM generating task plans by calling predefined action skills, which supports the planning part of the claim. Section 3.1 reports task-success results, but the paper does not establish that breaking long tasks into short, fixed movements increases success rates; its predefined skills are not necessarily short movements. (LLM, conf=0.86)
- Not matchedcites Chen et al., 2023b
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
plainstates a cause- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Partly supported
A result the paper reports from its own data — it cites no other source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“task execution efficiency is improved by 3.4× in composite tasks”
plainstates a cause- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
A result the paper reports from its own data — it cites no other source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“and 2.3× higher in complex tasks.”
plainstates a cause- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
32 more claims
Create a free account to see all 68 claims and every source verdict.
- Probably not a substantial claim. 22
- It may state two separate claims as one. 15
- The quote only partly states this claim. 1
Self-check
Percentages, sums and reported differences re-computed from the paper’s own numbers.
Each contradiction between two passages listed here was confirmed by an independent check.
Rated major: it would touch a result — a value a result rests on, reported and recomputed p-values on opposite sides of the significance level, two passages that conflict with no sign which is right, or a cited work retracted before this paper was published.
In the paper
§ 1 Introduction · p. 2
“Using this metric, we propose replacing the least recently used (LRU) unit whenever a new unit needs to be incorporated into a full memory.”
§ 1 Introduction · p. 2
“We propose to use hit rate as the metric of evaluating the effectiveness of memory replacement mechanism and present to always replace the least frequently used unit with the new unit.”
Rated minor: it would not touch a result — a value no result rests on, a p-value that stays on the same side of the significance level, or a conflict that is clerical, resolved elsewhere in the paper, or outside its own findings.
In the paper
§ 4.1 Experimental Setup and Metrics · p. 7
“The new dataset, ALFWorld-R, consists of long-sequence tasks”
§ 4.4 Replacement Policy Evaluation · p. 8
“This indicates that, on the ALFRED-R dataset, a larger window segment size in the W-TinyLFU policy allows for more effective utilization of memory units.”
Citations
Cites
Language Models are Few-Shot Learners
Brown, T. B., Mann, B., Ryder, N., Subbiah, M., Kaplan, J., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., et al. (2020) · arXiv · doi:10.48550/arxiv.2005.14165 · Cited by 23,372
Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A. C., Lo, W.-Y., et al. (2023) · 2023 IEEE/CVF International Conference on Computer Vision (ICCV) · doi:10.1109/iccv51070.2023.00371 · Cited by 11,953 · cited 1×
“Listing 2: Full Skill API and Action CODE used in the prompts.”
Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks
Lewis, P., Perez, E., Piktus, A., Petroni, F., Karpukhin, V., Goyal, N., Küttler, H., Lewis, M., Yih, W.-T., Rocktäschel, T., et al. (2020) · arXiv · doi:10.48550/arxiv.2005.11401 · Cited by 3,101
Generative Agents: Interactive Simulacra of Human Behavior
Park, J. S., O'Brien, J., Cai, C. J., Morris, M. R., Liang, P., & Bernstein, M. S. (2023) · Proceedings of the 36th Annual ACM Symposium on User Interface Software and Technology · doi:10.1145/3586183.3606763 · Cited by 2,187 · cited 2×
“Historical actions (Park et al., 2023), thoughts (Liu et al., 2023), contexts (Liang et al., 2023; Packer et al., 2023) are explored.”
Generative Agents: Interactive Simulacra of Human Behavior
Park, J. S., O'Brien, J. C., Cai, C. J., Morris, M. R., Liang, P., & Bernstein, M. S. (2023) · arXiv · doi:10.48550/arxiv.2304.03442 · Cited by 2,173
(2022) · doi:10.15607/rss.2023.xix.025 · Cited by 916
RT-1: Robotics Transformer for Real-World Control at Scale
Brohan, A., Brown, N., Carbajal, J., Chebotar, Y., Dabis, J., Finn, C., Gopalakrishnan, K., Hausman, K., Herzog, A., Hsu, J., et al. (2022) · arXiv · doi:10.48550/arxiv.2212.06817 · Cited by 727
Do As I Can, Not As I Say: Grounding Language in Robotic Affordances
Ahn, M., Brohan, A., Brown, N., Chebotar, Y., Cortes, O., David, B., Finn, C., Fu, C., Gopalakrishnan, K., Hausman, K., et al. (2022) · arXiv · doi:10.48550/arxiv.2204.01691 · Cited by 624
AI2-THOR: An Interactive 3D Environment for Visual AI
Kolve, E., Mottaghi, R., Han, W., VanderBilt, E., Weihs, L., Herrasti, A., Deitke, M., Ehsani, K., Gordon, D., Zhu, Y., et al. (2017) · arXiv · doi:10.48550/arxiv.1712.05474 · Cited by 578 · cited 1×
“We use the widely-adopted AI2-THOR simulator(Kolve et al., 2017) for evaluation.”
ChatGPT for Robotics: Design Principles and Model Abilities
Vemprala, S. H., Bonatti, R., Bucker, A., & Kapoor, A. (2024) · IEEE Access · doi:10.1109/access.2024.3387941 · Cited by 570 · cited 1×
“…as its core planning module can effectively decompose long and complex tasks into multiple short and fixed movements (Choi et al., 2024; Sarch et al., 2024; Chen et al., 2023b; Vemprala et al., 2023; Rana et al., 2023; Brohan et al., 2022, 2023; Belkhale et al., 2024), increasing the success rate.”
Research works printed but not linked: 8
- Chen, 2023, Llm-state: Expandable state representation for long-horizon task planning in the open world, arXiv preprint arXiv
- Lewis, 2020, Retrievalaugmented generation for knowledge-intensive nlp tasks, Advances in Neural Information Processing Systems, 33, 9459
- Li, 2023, Metaagents: Simulating interactions of human behaviors for llm-based task-oriented coordination via collaborative generative agents, arXiv preprint arXiv
- Liang, 2023, Unleashing infinite-length input capacity for large-scale language models with self-controlled memory system, arXiv preprint arXiv
- Chen, 2023, Gamegpt: Multiagent collaborative framework for game development, arXiv preprint arXiv
- Wang, 2023, c. when large language model based agent meets user behavior analysis: A novel user simulation paradigm, arXiv preprint arXiv
- Zhao, 2024, Chemdfm: Dialogue foundation model for chemistry, arXiv preprint arXiv
- 10.18653/v1/2023.findings-emnlp.226 · DOI · Paper page
Cited by
Enhancing reliability in LLM-integrated robotic systems: A unified approach to security and safety
Zhang, W. J., Kong, X., Dewitt, C., Bräunl, T., & Hong, J. B. (2025) · Journal of Systems and Software · doi:10.1016/j.jss.2025.112614 · Cited by 12
“These architectures can retain spatial layouts, past decisions, and failure states, enabling more consistent planning and better adaptation to complex or evolving environments [3, 35].”
3DLLM-Mem: Long-Term Spatial-Temporal Memory for Embodied 3D Large Language Model
Hu, W., Hong, Y., Wang, Y., Gao, L., Wei, Z., Yao, X., Peng, N., Bitton, Y., Szpektor, I., & Chang, K.-W. (2025) · Advances in Neural Information Processing Systems 38 · doi:10.52202/085713-2278 · Cited by 27
Dadu‐E : Rethinking the Role of Large Language Model in Robotic Computing Pipelines
Sun, W., Hou, S., Wang, Z., Bo, Y., Liu, S., Yang, X., Liang, S., Gan, Y., & Han, Y. (2025) · Journal of Field Robotics · doi:10.1002/rob.70120 · Cited by 11
“More details can be found in the experimental analysis and conclusions of this paper (Wang et al. 2024).”
Self-citationMemory in the Age of AI Agents
Hu, Y., Liu, S., Yue, Y., Zhang, G., Liu, B., Zhu, F., Lin, J., Guo, H., Dou, S., Xi, Z., et al. (2025) · arXiv · doi:10.48550/arxiv.2512.13564 · Cited by 389
Self-citationYu, Y., Yao, L., Xie, Y., Tan, Q., Feng, J., Li, Y., & Wu, L. (2026) · Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) · doi:10.18653/v1/2026.acl-long.981 · Cited by 4
Metadata
- Authors
- Wang, Z., Yu, B., Zhao, J., Sun, W., Hou, S., Liang, S., Hu, X., Han, Y., & Gan, Y.
- Affiliations
- Chinese Academy of Sciences (CN)
- Shenzhen Institute of Artificial Intelligence and Robotics for Society
- Alibaba Group (China) (CN)
- Beijing Institute of Technology (CN)
- Venue
- 2025 IEEE International Conference on Robotics and Automation (ICRA)
- Publisher
- IEEE
- Published
- 2025-05-19
- Pages
- 1-8
- ISSN
- 1050-4729 (ISSN-L)
- OpenAlex
- W4413945133
- Cited by
- 11
- References
- 32
- Citations by year
- 2026: 4 · 2025: 7
- FWCI
- 10.43
- Type
- conference paper
- Language
- en
- Open access
- green
- arxiv.org
- Locations
- Proceedings - IEEE International Conference on Robotics and Automation/Proceedings (publishedVersion)
- arXiv (Cornell University) (submittedVersion)
- arXiv (Cornell University)
- Funders
- National Key R&D Program of China (2024YFB4505800)
- National Natural Science Foundation of China (62025404)
- Beijing Municipal Science and Technology Commission (Z241100004224015)
- Topics
- Physical Sciences › Computer Science › Artificial Intelligence › AI-based Problem Solving and Planning
- Physical Sciences › Engineering › Control and Systems Engineering › Robot Manipulation and Learning
- Physical Sciences › Computer Science › Artificial Intelligence › Reinforcement Learning in Robotics
- Keywords
- embodied agents, long-term memory, short-term memory, 3D scene graphs, AI2-THOR, mobile manipulators
About these checks
Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.
The claim verdicts this report counts are experimental: their accuracy is not yet measured.
A language model drew the claims from the paper on 2026-10-29, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).
A second model labelled each claim on 2026-10-29 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).
Cited sources are checked with a language model’s help (gpt-6-sol).
Statistics and arithmetic are recomputed, not judged.
Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.
What counts as a claim ›How we measure confidence ›How we check this ›
Are you an author, or do you see a problem with this page? Request a review
How to cite
Cite CoChat for the checks on this page, and the paper for its own findings.
- This page
- CoChat. (2026). Citation check: KARMA: Augmenting Embodied AI Agents with Long-and-Short Term Memory Systems. https://app.cochat.ai/papers/10.1109/icra55743.2025.11128047
- The paper
- Wang, Z., Yu, B., Zhao, J., Sun, W., Hou, S., Liang, S., Hu, X., Han, Y., & Gan, Y. (2025). KARMA: Augmenting Embodied AI Agents with Long-and-Short Term Memory Systems. 2025 IEEE International Conference on Robotics and Automation (ICRA). https://doi.org/10.1109/icra55743.2025.11128047
Data sources
Bibliographic data: Semantic Scholar, OpenAlex, Crossref.
Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).
Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.