{"id":"db96fc1b-d070-420e-a5f7-66fd03d937fc","arxiv_id":"2608.04933","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"Separating world memory from task memory and grounding each goal in recalled evidence improves embodied-agent success rates by up to 42.5 points across 13 vision-language backbones.","lead":"Mimir is a memory system for embodied agents that tracks the state of the world separately from the progress of the task, and links the current goal to remembered objects before each action. On standard embodied benchmarks it raises success rates by up to 42.5 points across 13 different vision-language backbones.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"World-memory gains may hinge on simulator ground-truth state deltas rather than the memory architecture itself.","rationale":"The reader's weakest_assumption identifies exactly the load-bearing issue: the environment provides a structured state-change signal y_t with Delta_t, and world memory is updated with this signal through Eq. 8. My read of the full text confirms that this is the least secure condition for the central claim. The paper presents a coherent architecture, broad evaluation across 13 backbones, and an honest failure analysis, so this is not an internal inconsistency. But the scale of the claimed gains, especially the 12.5 EB-Habitat floor without world memory, makes the provenance of Delta_t decisive. If Delta_t is simulator ground truth, the headline result demonstrates a memory architecture that is partially fed privileged state, which is a weaker result for physical embodied agents. The concrete test I propose removes only the action-supported Delta_t from world-memory updates while preserving observation-based updates, directly isolating the assumption. Since the reader already reached CONDITIONAL with the same concern, my stress-test does not move the verdict; it reinforces the condition. I deliberately avoided treating disagreement with any external consensus as a flaw, and I credit the paper's broad backbone sweep and failure analysis as genuine supporting evidence. The recommended verdict remains UNCHANGED relative to the reader's CONDITIONAL assessment, with the explicit condition being disclosure and testing of the feedback assumption.","tokens_in":13466,"tokens_out":4031,"duration_ms":57068,"concrete_test":"Run Mimir on EB-Habitat with Qwen3-VL-8B and Qwen3-VL-32B under a modified signal: force Delta_t = empty for the world-memory action update U_act_W in Eq. 8, so W_{t+1} is updated only by U_obs_W from RGB-D observations, while keeping task-memory updates as in the original method. Compare EB-Habitat SR to Full Mimir (65.0/71.5) and to the no-world-memory rows (12.5/12.5) in Table 4. If SR stays near Full Mimir, the privileged delta is not the driver; if it falls toward 12.5, the central claim cannot be separated from the environment-feedback assumption.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that the architecture, not backbone capacity, drives the reported gains rests heavily on the world-memory component, whose ablation drops EB-Habitat SR from 65.0/71.5 to 12.5/12.5 (Table 4). However, the world-memory action update in Eq. 8 consumes y_t = (sigma_t, Delta_t), where Delta_t is described as 'feedback-supported change in object relation, object state, or hand state.' The paper never states how Delta_t is obtained in EB-ALFRED or EB-Habitat. If Delta_t is read from environment-internal state, Mimir is not inferring scene belief from perception; it is copying privileged simulator telemetry into symbolic memory. The 'Without world memory' ablation then measures the value of ground-truth state deltas, not the value of a learned or inferred world representation. This is a correctness-risk concern about transfer to physical embodied agents: a physical robot would have to estimate object relations and state changes from vision or interaction. The paper's failure analysis says memory 'cannot reconstruct evidence that never entered the perceptual stream,' which implicitly confirms that perception is the only legitimate source, yet the method section leaves the Delta_t source ambiguous. Without disclosure or an ablation isolating Delta_t, the load-bearing condition remains unverified.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Mimir, a neuro-symbolic memory system for long-horizon embodied manipulation. Mimir maintains two separable memories: a world memory that stores scene entities, relations, states, and evidence, and a task memory that tracks an ordered goal agenda, progress, hand state, failures, and constraints. Before each action, a grounding module binds the active goal to recalled world candidates, resolves missing source locations, and attaches evidence before planning. The system is evaluated on EB-ALFRED and EB-Habitat across 13 multimodal backbones, reporting consistent success-rate and goal-condition gains, with maximum gains of 42.5 and 37.5 percentage points on the two benchmarks, and a 86.0% success rate on the EB-Habitat Long-horizon subset using a 32B open-source backbone. The paper claims that the architecture, rather than backbone capacity, is the primary driver of the improvements, supported by ablations and a matched-backbone comparison against prior agent and memory systems.","tokens_in":13688,"tokens_out":3666,"duration_ms":48347,"significance":"If the empirical claim holds, the contribution is valuable: the explicit separation of world memory from task memory and the dynamic grounding operation provide a transparent interface between scene belief and execution progress, addressing a recognized bottleneck in long-horizon embodied agents. The breadth of the evaluation is a genuine strength: testing 13 backbones of varying scale and reporting uniformly positive deltas is stronger evidence than a single-model comparison. The failure analysis is also honest and useful, identifying cases where memory cannot compensate for ambiguous or missing perception. However, the significance depends on resolving a load-bearing ambiguity about the source of the environment feedback signal used to update world memory, and on clarifying the evaluation protocol with respect to variance and comparison subsets.","major_comments":[{"comment":"The source of y_t = (sigma_t, Delta_t) is never disclosed. In Eq. (8), the action update writes 'only effects supported by Delta_t', and the text describes Delta_t as 'feedback-supported change in object relation, object state, or hand state'. If Delta_t is obtained from simulator ground truth (e.g., the environment's internal object-relation and hand-state fields), then the world-memory component is copying privileged telemetry rather than inferring scene belief from perception. In that case, the 'Without world memory' ablation in Table 4, which drops EB-Habitat SR from 65.0 to 12.5 (Qwen3-VL-8B) and from 71.5 to 12.5 (Qwen3-VL-32B), measures the value of a privileged state-change signal, not the value of the proposed memory architecture per se. The manuscript must specify exactly how Delta_t is computed in EB-ALFRED and EB-Habitat, and should include an additional ablation where Delta_t is either estimated from observations only or removed entirely, so that the contribution of the grounding and memory structure can be separated from the contribution of the privileged signal.","section":"Method, Eq. (4) and Eq. (8); Ablation, Table 4"},{"comment":"The evaluation protocol is under-specified with respect to variance. The Metrics section states 'Each task is executed once' and then says the experiments were 'repeated ... and consistently obtained the same results', which is internally ambiguous: either the benchmark is deterministic given temperature 0, or the repetition is not described precisely. Because the central claim is that Mimir 'consistently improves' across all 13 backbones, the paper should state explicitly whether the environment is deterministic, report the number of seeds or runs if any, and provide a per-task breakdown of the four tasks that make up the Table 1 averages (Base, Common-sense, Complex-instruction, Long-horizon). Without this, the reported deltas are point estimates with no measure of variability.","section":"Experiments, Metrics and Table 1"},{"comment":"Table 2 is not directly comparable with Table 1. The caption says each value is the 'unweighted mean of Base and Long-horizon', following RoboMemory, whereas Table 1 reports four-task averages over Base, Common-sense, Complex-instruction, and Long-horizon. The claim that Mimir improves the overall average SR by 8.5% over prior systems is therefore based on a different subset than the headline numbers in Table 1. The comparison should be repeated on a consistent subset: either report the full four-task averages for all methods, or clearly state that the 8.5% improvement refers only to the Base+Long-horizon subset and confirm that the advantage persists on the other two tasks.","section":"Experiments, Table 2"}],"minor_comments":[{"comment":"There is a typo in the first sentence: 'Long-horizonembodiedtaskrequires' should be 'Long-horizon embodied tasks require'.","section":"Abstract"},{"comment":"The sentence 'And werepeated the experiments multipletimes and consistently obtained the same results' has grammatical errors and should be rewritten; it should also specify whether the repetition refers to multiple episodes of the same task or to re-running with different random seeds.","section":"Metrics"},{"comment":"The candidate list size K_t in Eq. (9) is a free parameter, but no sensitivity analysis is reported. A short discussion of how K_t is chosen and whether results are stable to its value would improve reproducibility.","section":"Method, Eq. (9)"},{"comment":"The phrase 'With a open-source backbone' should be 'With an open-source backbone'. In addition, the comparison with closed-source models would be clearer if the inference budget or decoding configuration of the closed-source baselines were specified, since the paper states only the maximum generation budget for all entries.","section":"Experiments, Table 3"}],"recommendation":"major_revision","confidential_remarks":"The main decision risk is the unverified source of the Delta_t signal. If the authors cannot disclose whether Delta_t is simulator ground truth or a perceptual estimate, the central claim about the architecture driving the gains is not yet convincing. I would encourage the editor to require this clarification and the corresponding ablation before further consideration, because the current setup could be measuring the value of privileged state access rather than the value of the proposed memory system."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe thing to know: Mimir is a genuinely reusable memory wrapper for long-horizon embodied agents, and the core idea—dynamically grounding the active goal against a separate world memory before each action—is presented clearly and tested more broadly than most memory papers. The 13-backbone comparison is the real strength; the gains are consistent and large, and the 86.0% EB-Habitat Long-horizon result against closed-source models is attention-worthy if it holds up.\n\nWhat's new: prior work like LLM-State, CAPEAM, and RoboMemory has pieces of world and task memory, but Mimir adds an explicit grounding operator that binds the active goal to recalled candidates, fills missing source locations from world memory, and fails closed when no evidence-supported binding exists. That interface is the contribution, and the ablation structure (world memory vs task memory) does roughly what you'd expect.\n\nThe soft spot is load-bearing. The world-memory action update (Eq. 8) consumes y_t = (sigma_t, Delta_t), where Delta_t is described as \"feedback-supported change in object relation, object state, or hand state.\" The paper never discloses where Delta_t comes from. If it's read from simulator internal state, then the \"world memory\" is partly copying privileged telemetry into symbolic memory, not inferring scene belief from vision. The ablation in Table 4 shows EB-Habitat SR collapses from 65.0 to 12.5 without world memory, so the entire empirical case for the architecture rests on this signal. On a physical robot, the agent would have to estimate these deltas from raw perception, which is a much harder problem. The failure analysis's honest line that memory \"cannot reconstruct evidence that never entered the perceptual stream\" makes the omission more puzzling.\n\nOther concerns are minor. There are no error bars or seed variation, though temperature=0 and multiple consistent runs soften that. Table 2's comparison to RoboMemory uses the Base+Long-horizon subset, not the four-task average in Table 1; it's labeled, so not deceptive, but easy to misread.\n\nBottom line: this paper is worth a serious referee. The architecture is coherent, the evaluation is broad, and the delta-source issue is fixable by disclosure and an ablation that replaces privileged deltas with vision-only state estimation. If the authors confirm the source and release code, the result could be an important subfield contribution. Send it to review.","headline":"Mimir's memory architecture is solid and broadly evaluated, but the main gains may hinge on an undisclosed simulator feedback signal.","tokens_in":14220,"tokens_out":2921,"would_cite":false,"duration_ms":32971,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Mimir claims that separating world memory from task memory, and dynamically grounding each active goal in recalled world evidence before acting, is what drives long-horizon embodied success; across 13 vision-language backbones it reports…","keywords":["embodied agents","long-horizon manipulation","neuro-symbolic memory","dynamic grounding","world memory","task memory","vision-language models","EB-Habitat"],"falsifier":"Run Mimir on EB-Habitat with the $\\Delta_t$ channel disabled so world memory is updated only from raw observations, and compare success rate with the full system. If it collapses toward the 12.5% of the 'without world memory' ablation, the reported gains depend on privileged state feedback rather than on the memory organization itself.","tokens_in":13268,"feed_emoji":"🤖","tokens_out":11293,"duration_ms":114560,"temperature":0.7,"pith_summary":"The paper argues that long-horizon embodied agents fail not because they forget too much, but because their memory does not expose an actionable state for the goal currently being executed. Mimir therefore separates world memory, which tracks where objects are and what state they are in, from task memory, which tracks what remains undone and what has failed. Before each action, a grounding step binds the active goal to the best-supported candidate recalled from world memory, fills missing source locations, and attaches the evidence behind the binding. On EB-ALFRED and EB-Habitat, the paper reports that this design improves success rate for all 13 tested backbones, with gains up to 42.5 percentage points and a mean gain of 23.0 points, and that a 32B open model reaches 86.0% success on the EB-Habitat Long-horizon subset, above every closed-source model tested. If true, the explicit task–world interface, not backbone scale, is what makes long-horizon performance possible.","feed_headline":"23-point gain: split world and task memory lifts embodied agents","feed_subtitle":"Separating world state from task progress lets a 32B model beat closed-source models on long-horizon tasks.","key_machinery":"The load-bearing mechanism is the closed loop $(T_t,W_t) \\to$ grounding $\\to$ abstract action $\\to$ embodiment adapter $\\to$ execution feedback, with the grounding operator $\\hat{g}_t = G(g_t,R_t,F_t,A_e,x,o_t)$ as the decisive step. Recall returns a bounded candidate list from world memory with each candidate's parent and an evidence score; grounding filters candidates that violate constraints or are contradicted by failures, selects the best-evidence candidate, fills an omitted source from the remembered parent, and returns $\\bot_{\\mathrm{grd}}$ rather than an unsupported binding. That fail-closed property forces the agent to explore or block a goal instead of acting on an ungrounded query. The two memory write paths are what make the loop stable: observation updates add or refine hypotheses, action updates only write effects supported by $\\Delta_t$, and task memory's explicit goal-status map $s_t$ decides completion by whether $\\Delta_t$ establishes the goal postcondition.","core_discovery":"Mimir's central claim is that memory for embodied agents should be modeled as two evolving states—a world memory $W_t$ holding an episode-level tree of entities, parent relations, attributes, and evidence, and a task memory $T_t$ holding an ordered goal agenda, per-goal status, hand state, failed bindings, and completion evidence. The two are coupled by dynamic grounding: task memory exposes the next unfinished goal, goal-conditioned recall $R(W_t^+, Q(g_t))$ returns a bounded list of candidate entities with their remembered parents and supporting evidence, and the grounding operator $G$ commits to one admissible candidate, resolves a missing source from that candidate's parent, and attaches evidence before planning and execution. World memory is updated by distinct observation and action operators, and the action operator writes only changes supported by the environment feedback $y_t=(\\sigma_t,\\Delta_t)$, where $\\Delta_t$ is the feedback-supported change in object relation, object state, or hand state. The paper claims that this separation and grounding, evaluated over thirteen multimodal backbones on EB-ALFRED and EB-Habitat, improves success rate and goal-condition success rate in every case, with the largest gains exactly where the bare backbone is weakest, and that the resulting system outperforms prior agent and memory systems under a matched backbone and surpasses closed-source foundation models on long-horizon tasks.","pith_inferences":["Editorial inference: the dependence on the structured feedback signal $\\Delta_t$ is the main transfer risk; a direct test would disable or corrupt that signal and measure how much of the world-memory gain survives.","Editorial inference: since world memory is reset every episode, extending Mimir with persistent cross-episode semantic facts is a natural next step that the paper does not explore.","Editorial inference: the fail-closed grounding behavior suggests a cheap diagnostic for any embodied planner: count decisions that act without an evidence-supported binding; the paper's account predicts those actions are a major source of failure in long-horizon tasks."],"forward_implications":["With Mimir, InternVL3-8B reaches EB-ALFRED SR 60.0 and GC 68.5, roughly matching or exceeding its 38B and 78B counterparts, so structured memory can substitute for some backbone scale.","Removing world memory drops EB-Habitat SR from 65.0 to 12.5 with Qwen3-VL-8B and from 71.5 to 12.5 with Qwen3-VL-32B, so the scene-belief store is the source of the largest single-component gain in interactive environments.","Removing task memory costs 23.0 SR points on Qwen3-VL-8B and 37.5 points on Qwen3-VL-32B on EB-ALFRED, showing that an ordered goal agenda carries much of the instruction-following benefit.","Under the same Qwen2.5-VL-72B backbone, Mimir raises EB-Habitat SR by 16.0 points and GC by 13.6 points over the strongest prior memory system, RoboMemory.","Instruct mode with Mimir outperforms Thinking mode with Mimir on all four metrics, suggesting that externalized progress and evidence can substitute for additional internal reasoning in these tasks."],"supporting_citations":[{"why":"Defines the EmbodiedBench suite with EB-ALFRED and EB-Habitat, supplying every task, metric, and backbone-evaluation protocol the experiments use.","marker":"Yang et al. 2025"},{"why":"RoboMemory is the strongest prior memory system compared under the same Qwen2.5-VL-72B backbone, and the matched-backbone improvement is the paper's headline comparison.","marker":"Lei et al. 2025a"},{"why":"LLM-State provides the open-world state-tracking baseline that motivates the separation of world belief from task progress.","marker":"Chen, Xiao, and Hsu 2023"},{"why":"CAPEAM supplies the context-aware planning and environment-aware memory baseline that Mimir's explicit write and grounding interface is contrasted with.","marker":"Kim et al. 2023"},{"why":"Voyager is one of the generic agent frameworks in the matched-backbone comparison table.","marker":"Wang et al. 2023"},{"why":"Reflexion is one of the generic agent frameworks in the matched-backbone comparison table.","marker":"Shinn et al. 2023"},{"why":"Cradle is one of the generic agent frameworks in the matched-backbone comparison table.","marker":"Tan et al. 2024"},{"why":"RoboOS is one of the generic agent frameworks in the matched-backbone comparison table.","marker":"Tan et al. 2025"}],"fun_headline_variants":["Mimir lifts embodied agents 23% on average with dual-memory grounding","World and task memory split yields 23% average gain on embodied tasks","Neuro-symbolic memory dynamic grounding hits 86% on long-horizon","Max gain 42.5%: Mimir's memory separation improves all backbones"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the environment supplies a structured state-change signal, $\\Delta_t$, stating exactly which object relation, object state, or hand state changed after each action, and that world memory is updated only from this signal. A physical robot would have to recover that signal from raw vision, and the large world-memory gains rest on it.","fun_headline_variants_meta":{"raw":{"variants":["Mimir lifts embodied agents 23% on average with dual-memory grounding","World and task memory split yields 23% average gain on embodied tasks","Neuro-symbolic memory dynamic grounding hits 86% on long-horizon","Max gain 42.5%: Mimir's memory separation improves all backbones"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000274,"raw_usage":{"total_tokens":1698,"prompt_tokens":1060,"completion_tokens":638,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":676,"completion_tokens_details":{"reasoning_tokens":553}},"tokens_in":676,"tokens_out":638,"duration_ms":7374,"temperature":1.0,"reasoning_tokens":553,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T13:33:19.476910+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run Mimir on EB-Habitat with the $\\Delta_t$ channel disabled so world memory is updated only from raw observations, and compare success rate with the full system. If it collapses toward the 12.5% of the 'without world memory' ablation, the reported gains depend on privileged state feedback rather than on the memory organization itself.","supporting_citations":[{"cited_title":"2023 , eprint =","cited_arxiv_id":null,"evidence_quote":"Voyager is one of the generic agent frameworks in the matched-backbone comparison table."},{"cited_title":"2023 , eprint =","cited_arxiv_id":null,"evidence_quote":"Reflexion is one of the generic agent frameworks in the matched-backbone comparison table."}],"review_version":1}