Recent studies of embodied agents based on frontier multimodal large language models (MLLMs) develop harness engineering around the MLLM, such as harnesses with post-trained policies and recursive self-improvement loops. Yet whether these additions exceed the native capability of the MLLM remains unexplored. We introduce SimplEvo to explore how far an MLLM can improve under direct control when it retains its execution experience in self-written text memories. It requires no weight updates, post-trained policies, or privileged information. The MLLM writes a coordinate-free task memory once from a single demonstration. Under a strict train-test protocol, it then creates a bounded working memory on training episodes only and refines it after each episode with lessons on procedure, completion, and recovery. Both memories stay fixed at test time. On EBench, SimplEvo lifts the success rate (SR) of GPT-6-Astra under direct control from 46.7% to 81.7%, and tops the RoboDojo leaderboard with 69.63% SR and a score of 75.32. Compared with memories that also store past visual frames, SimplEvo reaches a similar SR with text memory alone while using fewer tokens. With its memories fixed, SimplEvo also generalizes to unseen layouts and randomized scene configurations. Moreover, the working memory transfers across models: written by GPT-6-Astra, it can be reused as is by weaker MLLMs and improve their performance without any adaptation of their own.
Figures & tables
Figure 1: GPT-6-Astra in two baseline configurations, direct control and a harness with policy, and with SimplEvo, whose working memory is written on training episodes and then fixed. (a) Each configuration with its EBench success rate (SR) and InternW0- Δ as the harness policy. The bar under each shows which component commands the robot, and the MLLM does not see the actions of a delegated policy. (b) Overall SR on EBench and (c) on RoboDojo, where † marks the strongest post-trained policy and zero-shot denotes the official entry of GPT-6-Astra. (d) SR on the 12 RoboDojo Generalization tasks with standard and randomized configurations.
Figure 2: Overview of SimplEvo. During online adaptation (left), an annotator instance of GPT-6-Astra reviews one training demonstration and writes a coordinate-free task memory of the stages and the completed state, and after every training episode, GPT-6-Astra rewrites the working memory from the trajectory and outcome of that episode into lessons on procedure, completion, and recovery. Once fixed (right), both memories are read in new scenes, and other MLLMs read the working memory of GPT-6-Astra without modification. All panels show peg insertion on EBench.
Overall
Precision
Mobility
Horizon
System
SR
Score
Low
Med.
High
Mobile
Tabletop
Short
Long
Post-trained policies
X-VLA [ Zheng et al., 2025 ]
23.7
35.8
35.7
13.3
2.5
29.3
8.6
45.9
7.4
Fast-WAM [ Yuan et al., 2026b ]
25.6
37.1
35.7
19.0
3.8
34.0
2.9
46.4
10.4
Cosmos3-Edge [ NVIDIA, 2026 ]
29.3
42.5
38.6
20.8
13.8
33.2
18.6
39.1
22.1
GigaBrain-0.7 [ GigaBrain Team et al., 2026 ]
33.3
46.0
35.4
33.8
25.0
31.6
37.9
40.0
28.3
Table 1: SR (%) and score on the 510 test episodes of EBench for the post-trained policies on the public leaderboard (29 October 2026) and for GPT-6-Astra without working memory and with SimplEvo. Bold and underline mark the best and second-best values per column.
Overall
w/o Open
Capability SR (%)
System
Class
SR
Score
SR
Score
Gen.
Prec.
Long
Mem.
Open
Post-trained policies
π0.5 [ Black et al., 2025a ]
VLA
6.93
11.44
8.25
13.80
8.17
5.50
14.67
4.67
1.67
Awomo-0.5 [ Awomo-WAM Team, 2026 ]
WAM
29.64
35.34
31.33
38.15
23.84
33.50
26.75
41.22
22.92
VPP2 [ Guo et al., 2026b ]
WAM
32.26
39.26
39.31
47.93
29.50
35.17
37.00
55.56
4.08
WALL-OSS-1.0-lite [ Zhai et al., 2025 ]
VLA
34.97
42.59
38.82
48.05
37.94
29.08
32.25
56.00
19.58
Table 2: SR (%) and score on the 42 tasks of RoboDojo, over 50 layouts per task, for entries on the public leaderboard (2 November 2026) and for SimplEvo on its test layouts. The w/o Open columns average the four dimensions other than Open, as on the leaderboard. Bold and underline mark the best and second-best values.
Overall
w/o Open
Open
Configuration
Layouts (Episodes)
SR
Score
SR
Score
SR
Task memory only
15 training (630)
36.37
43.69
37.34
45.78
32.50
SimplEvo, online
15 training (630)
58.53
65.04
55.66
63.15
70.00
SimplEvo, fixed working memory
50 test (2,100)
69.63
75.32
67.67
74.09
77.50
Table 3: SR (%) and score of GPT-6-Astra on the 42 tasks of RoboDojo: task memory only and online adaptation on the 15 training layouts per task, and the fixed working memory of SimplEvo on the 50 test layouts per task. Bold marks the best value in each column.
Memory
Coffee
Gear
Peg
Nut
SR
Score
Tokens
Minutes
Task memory only
0.0
40.0
10.0
10.0
15.0
40.5
3.51
29.5
+ text working memory (SimplEvo)
20.0
70.0
90.0
10.0
47.5
64.1
2.86
23.0
Text working memory + visual replay
30.0
70.0
70.0
20.0
47.5
64.5
3.47
27.2
VISTA history [ Han et al., 2026 ]
20.0
90.0
40.0
30.0
45.0
68.0
6.10
31.4
Table 4: Memory designs of GPT-6-Astra under direct control on one online chain of ten validation scenes per high-precision EBench task: SR (%) per task and overall, mean score, tokens per episode (M), and minutes of control per episode at equal simulator speed (model time plus 0.5 s per physics step). The visual-replay row includes text working memory, and VISTA history keeps the whole chain in one conversation. Bold marks the best value in each column.
Memory
Coffee
Gear
Peg
Nut
SR
Score
Tokens
Minutes
Harness with π0.5
Task memory only
0.0
60.0
10.0
20.0
22.5
41.3
1.73
27.9
+ text working memory
0.0
60.0
30.0
10.0
25.0
51.6
2.88
26.5
Text working memory + visual replay
10.0
20.0
10.0
10.0
12.5
40.7
2.57
27.0
Text + π0.5 trajectory
10.0
40.0
0.0
20.0
17.5
43.2
2.78
28.0
Direct control
Table 5: Working memory under the two agent harnesses of GPT-6-Astra on the chains and with the columns of Table 4 . The visual-replay row includes text working memory, and the trajectory row extends it with the executed motion of π0.5 , as Appendix E.3 describes. Bold marks the best value within each harness.
Figure 3: Test episodes with the fixed working memory of SimplEvo (green) and without working memory on the same scenes, with verbatim excerpts of the memory read in each SimplEvo episode ([…] marks omissions). Each row shows a success of SimplEvo on a scene where the other configurations fail. Over all test episodes of these tasks, SimplEvo succeeds in 18 of 20 in (a), 12 of 20 in (b), and 49 of the 50 test layouts in (c), compared with none in (a) and 2 in (b) for direct control without working memory, and none and 3 for the harness with π0.5 . Dark labels give commanded positions in meters equal to coordinates in the working memory, which GPT-6-Astra wrote from its own commands and proprioception in training episodes without access to ground-truth object positions. Baseline frames in (a) and (b) come from Table 1 and in (c) from Appendix E.5 .
System
Class
Seen / standard
Unseen / randomized
Δ
EBench, 15 multi-layout tasks: seen and unseen layouts
Direct control [ Yao et al., 2026 ]
Agent
73.8
63.1
− 10.7
Harness with π0.5 [ Yao et al., 2026 ]
Agent + VLA
77.2
69.6
− 7.6
Harness with InternW0- Δ [ Yao et al., 2026 ]
Agent + WAM
74.5
71.7
− 2.8
SimplEvo (ours)
Agent
83.1
87.5
+ 4.4
RoboDojo, 12 Generalization tasks: standard and randomized configurations
Table 6: SR (%) by layout overlap on EBench and scene configuration on RoboDojo. Seen EBench layouts occur in at least one completed training episode of SimplEvo. The same partition applies to every configuration of GPT-6-Astra. On RoboDojo, each Generalization task pairs a standard configuration with a randomized one, the other rows are the entries of the public leaderboard (2 November 2026), and the 15 training layouts of SimplEvo include standard and randomized configurations. SR is averaged within each task and then across tasks, and Δ is the change from the first to the second column. Class marks vision-language-action models (VLA), world action models (WAM), and MLLM agents (Agent), and Agent + VLA or Agent + WAM denotes an agent that calls such a policy. Bold and underline mark the best and second-best values within each benchmark.
Task memory only
Task + working memory of GPT-6-Astra
Model
SR
Score
SR
Score
Other MLLMs
Gemini 3.8 Flash
23.8
37.8
53.8
71.7
DeepSeek-V4.1-Flash
0.8
4.6
13.1
26.5
MLLM that wrote the working memory
GPT-6-Astra
60.8
76.8
81.5
89.3
Table 7: Transfer of the working memory of GPT-6-Astra across MLLMs, measured on the 130 test episodes of the Mixed generalization setting of EBench, five per task across all 26 tasks: SR (%) and mean score with task memory only and with the working memory of GPT-6-Astra read without modification. The working-memory entry of GPT-6-Astra comes from Table 1 . Bold marks the higher value for each model.
Appendix figures & tables6 assets
Supplementary material from the paper’s appendix.
Appendix
Figure B.3: Adaptation chains of SimplEvo. (a) SR at each chain position of the online RoboDojo chains (42 episodes per point, averaged over tasks without dimension weighting) with working memory and with task memory only. (b) Mean size of the working memory read at each position, over the 42 RoboDojo chains and over the EBench adaptation chains still running at that position (26 at positions 1 to 3, 5 at positions 14 and 15).
System
Toast
Mugs
Sweep
Plug
Swap
Kong
Pour
Mean
RAX, ≤ 16
2.5
36.9
53.3
94.0
100.0
64.0
100.0
64.4
RAX, ≤ 64
47.2
39.2
71.7
94.0
100.0
96.0
100.0
78.3
SimplEvo, 15
26.0
24.0
56.0
80.0
100.0
96.0
98.0
68.6
Appendix
Table D.2: Score (0–100) on the seven RoboDojo tasks evaluated by RAX [ Huang et al., 2026 ] , with the first three tasks in their randomized configurations. Italic rows come from the RAX paper with at most 16 or 64 exploration episodes per task, and its programs run on metric depth and camera poses outside our information protocol. SimplEvo uses the working memory fixed after 15 training episodes and is evaluated on its test layouts.
Task memory only
Task + working memory
Model
SR
Score
SR
Score
Put glasses into a glassbox
Gemini 3.8 Flash
0.0
15.0
90.0
91.7
MiMo-V2.6-Pro
0.0
6.7
70.0
78.3
DeepSeek-V4.1-Flash
0.0
6.7
40.0
45.0
GLM-5.3-Flash
0.0
3.3
20.0
25.0
Appendix
Table F.1: SR (%) and mean score of four MLLMs on two EBench tasks, 20 episodes each, with task memory only and with the working memory that GPT-6-Astra wrote at the same chain position.
Figure H.1: Coordinate replay on two high-precision single-layout tasks of EBench, peg insertion and gear installation. Both episodes of SimplEvo with the fixed working memory (green) succeed, and direct control and the harness with π0.5 fail on the same scenes. Dark labels indicate commanded positions in meters that match coordinates in the working memory, the cards quote the lines that store them, and […] marks internal omissions. Following Appendix E.6 , arm moves count the moving commanded waypoints of the episode, whose target lies more than 1 cm from the previous target of the same arm. Frames come from the wrist cameras, and the last frame shows the terminal state.
Figure H.2: Lessons on unseen layouts. (a, b) EBench test scenes of the fruit milkshake task and of the teacup and teapot task, whose layouts appear in no training episode of the chain. (c, d) RoboDojo test layouts of picking from a conveyor by image and of solving an equation. All four episodes of SimplEvo with the fixed working memory (green) succeed. On the same scenes, direct control and the harness with π0.5 fail in (a) and (b), and GPT-6-Astra with task memory only fails in (c) and (d). The cards quote lessons of the working memory read in the episode, and Agent marks a message of the agent in the same episode. Dark labels give commanded positions in meters that equal stored coordinates. EBench counts cover the moving commanded waypoints, and RoboDojo counts cover the arm targets away from the home pose. Frames come from the wrist cameras, and in (c) and (d) also from the head camera.
Figure H.3: Failures of SimplEvo with the fixed working memory on three EBench tasks: nut tightening (score 50.0), pens into a holder (score 66.7), and sandwich making (score 75.0). Each episode ends at the physics step limit of its task. The cards quote the working memory read in the episode, and Agent marks the last message of the agent in the episode. Frames come from the wrist cameras.
Memory allows large language model (LLM) agents to use past experience, often stored as textual records, when responding to requests. For embodied agents that control robots, visual and spatial observations introduce a broader choice of memory modalities, affecting both the information available for subsequent tasks and the model inference costs of constructing and using memory. We introduce MeMoREval, a benchmark for studying this problem through a controlled experience-request protocol. Agents first experience a sequence of events without knowing the subsequent tasks, and subsequently they use memory to respond to information and action requests without receiving new observations. The benchmark comprises 53 tasks across four simulated indoor environments and measures task success alongside the inference costs of memory construction and use. Our reference evaluations reveal that storing frames outperforms storing generated descriptions at lower write-time inference cost for the primary model, while storing frames with scene graphs achieves the highest aggregate task success for every tested model. Extensions to accumulated experience, selective frame retention, and visual input to memory writers demonstrate how the benchmark accommodates further memory-design investigations. Two physical-robot demonstrations establish the feasibility of applying its protocol beyond simulation.
Junxiang Wang, Yuchen Liu, Yi Wu +6
Dexmate · Carnegie Mellon University · Purdue University
Deploying Multimodal Large Language Models as the brain of embodied agents remains challenging, particularly under long-horizon observations and limited context budgets. Existing memory assisted methods often rely on textual summaries, which discard rich visual and spatial details and remain brittle in non-stationary environments. In this work, we propose a non-parametric memory framework that explicitly disentangles episodic and semantic memory for embodied exploration and question answering. Our retrieval-first, reasoning-assisted paradigm recalls episodic experiences via semantic similarity and verifies them through visual reasoning, enabling robust reuse of past observations without rigid geometric alignment. In parallel, we introduce a program-style rule extraction mechanism that converts experiences into structured, reusable semantic memory, facilitating cross-environment generalization. Extensive experiments demonstrate state-of-the-art performance on embodied question answering and exploration benchmarks, yielding a 7.3% gain in LLM-Match and an 11.4% gain in LLM MatchXSPL on A-EQA, as well as +7.7% success rate and +6.8% SPL on GOAT-Bench. Analyses reveal that our episodic memory primarily improves exploration efficiency, while semantic memory strengthens complex reasoning of embodied agents.
Ji Li, Bo Wang, Jing Xia +2
The University of Hong Kong, Hong Kong SAR. · Beijing Institute of Technology, Beijing, China.
Most Large Language Model (LLM) agent memory systems rely on a small set of static, hand-designed operations for extracting memory. These fixed procedures hard-code human priors about what to store and how to revise memory, making them rigid under diverse interaction patterns and inefficient on long histories. To this end, we present MemSkill, which reframes these operations as learnable and evolvable memory skills, structured and reusable routines for extracting, consolidating, and pruning information from interaction traces. Inspired by the design philosophy of agent skills, MemSkill employs a controller that learns to select a small set of relevant skills, paired with an LLM-based executor that produces skill-guided memories. Beyond learning skill selection, MemSkill introduces a designer that periodically reviews hard cases where selected skills yield incorrect or incomplete memories, and evolves the skill set by proposing refinements and new skills. Together, MemSkill forms a closed-loop procedure that improves both the skill-selection policy and the skill set itself. Experiments on LoCoMo, LongMemEval, HotpotQA, and ALFWorld demonstrate that MemSkill improves task performance over strong baselines and generalizes well across settings. Further analyses shed light on how skills evolve, offering insights toward more adaptive, self-evolving memory management for LLM agents. Code is available at https://github.com/ViktorAxelsen/MemSkill
Haozhen Zhang, Quanyu Long, Jianzhu Bao +4
Nanyang Technological University · University of Illinois Urbana-Champaign · University of Illinois Chicago +1