{"id":"e17a8db4-cbeb-4af2-bd25-02f9fd028796","arxiv_id":"2607.28156","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"Query-only reflective experience memory distilled from past retrieval trajectories improves cross-task evidence search in long-video multimodal agents without feeding historical facts into answers.","lead":"RRM adds a third memory layer that stores how to search long videos, not what happened in them, and uses that only to rewrite retrieval queries. It beats strong memory-agent baselines on three long-video benchmarks while cutting search rounds.","discovery_kind":"new_method","skeptic_critique":{"model":"grok-4.5","headline":"The central claim rests on unmeasured isolation of “procedural” experience; residual task-specific leakage could still drive the reported gains.","rationale":"The reader correctly located the load-bearing premise: procedural purity plus the mini-batch delayed-GT protocol are asserted by construction rather than measured. That premise is what lets the authors claim “reflective retrieval memory” (not merely a stronger Search–Answer controller or soft test-time adaptation) explains the controlled lifts over M3-Agent. Fig. 3a and Table 2 are supportive but downstream; they do not substitute for a leakage audit. No stronger internal inconsistency appears—the architecture is coherent and the M3-Agent-matched backbone is a fair control—so the verdict stays CONDITIONAL rather than moving to REJECT. A direct audit of experience records and guided-retrieval logs is the single check that would either solidify or undercut the central attribution. Reproducibility gaps and missing error bars remain secondary limits already noted by the reader.","tokens_in":13205,"tokens_out":662,"duration_ms":13618,"concrete_test":"Sample 50 stored experiences (25 success / 25 failure) from a finished M3-Bench-Web run; have an independent annotator (or a held-out LLM with a strict fact-detection rubric) flag any residual source-video entities, events, answers, or option labels. Separately, for 30 anomaly-triggered steps, log the experience-derived focus f_ref and the subsequent Search hits; measure overlap of retrieved entities/timestamps with the experience’s source video. If >10% of records or >15% of guided retrievals show leakage, the isolation premise fails and the headline attribution weakens.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The strongest claim requires that LLM-extracted experiences, after stripping answers/entities and enforcing cross-video isolation, act only as pure search-control signals, so that answer gains come solely from better current-video evidence (Three-Layer Memory Architecture; Triggered Query-Only Experience Reuse; mini-batch protocol). The paper asserts this via design (query-only injection, source-video exclusion, delayed batch feedback) and shows that query-only beats prompt-level injection (Fig. 3a). It never directly audits residual leakage: no inspection of extracted records for lingering entities/events/answers, no retrieval-log analysis of whether experience-guided queries still surface source-task facts, and no controlled “poisoned-experience” or same-domain vs. cross-domain transfer test. If the extractor or selector still injects task-specific cues that bias which current-video segments are fetched (or if co-batching same-video questions plus LLM judging softens the online protocol), the 9.1/5.8/7.4 gains and round reductions cannot be cleanly attributed to transferable procedural knowledge. Ablations and efficiency results remain consistent with a weaker “better retrieval controller” story that does not need the isolation premise.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.5","summary":"The paper proposes Reflective Retrieval Memory (RRM), which augments an entity-centric multimodal factual memory graph (episodic + semantic, following M3-Agent) with a third layer of reflective experience memory. From historical task trajectories, RRM distills structured procedural retrieval records (evidence requirements, query patterns, failure modes, adjustment strategies) into separate success and failure banks, reuses them only as query-level control signals when anomaly triggers fire, and keeps answer generation conditioned solely on newly retrieved current-video evidence. Online Query Reflection provides within-task repair; a lifecycle mechanism merges, decays, and prunes experiences. Under a mini-batch online protocol with delayed ground-truth feedback and cross-video isolation, RRM reports gains of 9.1, 5.8, and 7.4 points over M3-Agent on M3-Bench-Robot, M3-Bench-Web, and Video-MME-Long, with fewer average retrieval rounds, supported by progressive ablations and a prompt-injection vs. query-only comparison.","tokens_in":13484,"tokens_out":1318,"duration_ms":40928,"significance":"The work targets a genuine bottleneck in long-horizon multimodal agents: not only what to store, but how to improve retrieval strategy across tasks without contaminating current-video reasoning with historical facts. The separation of procedural experience from factual memory, the query-only reuse constraint, and the controlled comparison against the shared M3-Agent backbone are clear design strengths. Consistent accuracy gains across three benchmarks, complementary ablations (OQR, experience memory, lifecycle), the query-only vs. prompt-level result (Fig. 3a), and simultaneous reductions in retrieval rounds make a credible empirical case that retrieval-control learning is useful. If the isolation premise holds under stronger audits, the framework is a transferable template for memory-augmented multimodal agents beyond the specific backbone.","major_comments":[{"comment":"Three-Layer Memory Architecture and Triggered Query-Only Experience Reuse: the central attribution—that gains come from transferable procedural retrieval knowledge rather than residual task-specific cues—is enforced by design (answer/entity stripping, cross-video exclusion, query-only injection) and indirectly supported by Fig. 3(a), but never directly measured. There is no audit of extracted records for lingering entities/events/answers, no retrieval-log analysis of whether experience-guided queries still surface source-task content, and no controlled same-domain vs. cross-domain or poisoned-experience test. Without at least one such check, the 9.1/5.8/7.4 gains cannot be cleanly attributed to pure search-control transfer versus a stronger generic retrieval controller. A compact qualitative sample of records plus a simple leakage or transfer split would make the claim load-bearing rathe","section":"Three-Layer Memory Architecture; Triggered Query-Only Experience Reuse; Fig. 3(a)"},{"comment":"Failure-Aware Experience Modeling and the mini-batch protocol: failure experiences are extracted with access to y*_i and ŷ_i after the batch is scored. That is consistent with the stated delayed-feedback protocol, but it means the failure bank depends on ground-truth labels that are unavailable in true label-free deployment. The paper should either (i) report a success-only / no-GT ablation on the same tables, or (ii) clearly scope failure-bank updates as an offline or weakly supervised mode and show that main gains survive without M−. As written, Table 2’s “Reflective Experience Memory” conflates two regimes with different supervision assumptions.","section":"Failure-Aware Experience Modeling; Evaluation Protocol; Table 2"},{"comment":"Implementation Details / free parameters: several load-bearing knobs (mini-batch size 64 with same-video co-batching, max five Search–Answer rounds, at most one success and one failure experience per step, Dedup_L, lifecycle thresholds for merge/prune/cold-start/decay) are fixed without sensitivity analysis. Because experience write timing and anomaly-triggered reuse both depend on these choices, at least a short sensitivity or stability check on batch size and experience cap (e.g., on one benchmark) is needed to show that the headline margins are not brittle to the chosen online schedule.","section":"Implementation Details; Online Memory Lifecycle Management; Eq. (4)"}],"minor_comments":[{"comment":"Abstract and early pages show systematic missing word spaces (e.g., “mostmethodsemphasize,” “tasktrajectories,” “ReflectiveRetrievalMemory”). Clean the camera-ready text; if this is a PDF extraction artifact, verify the source.","section":"Abstract; Introduction"},{"comment":"Figure 2 is central but dense; a short caption walkthrough of the A/B/C path (OQR vs. experience reuse vs. lifecycle write-back) would help readers separate within-task repair from cross-task transfer.","section":"Figure 2"},{"comment":"Table 1: NS-Mem has no Video-MME-Long entry; state explicitly whether the method is inapplicable or was not run, to avoid an incomplete SOTA row.","section":"Table 1"},{"comment":"Eqs. (1)–(4) are clear, but Dedup_L and K^base_i,t are only lightly defined in-text; point to the appendix schema or give a one-line definition in the main text.","section":"Reflective Retrieval Control; Eq. (4)"},{"comment":"Related Work correctly positions RRM against Reflexion/ExpeL/AWM/ReMe; a single sentence on why multimodal long-video retrieval makes prompt-level experience injection especially risky (vs. text-only agents) would sharpen the contrast already tested in Fig. 3(a).","section":"Related Work"}],"recommendation":"major_revision","confidential_remarks":"Solid empirical systems paper with a clean idea and controlled baseline. The isolation premise is the main intellectual risk; if the authors add a leakage/transfer audit and a no-GT or success-only ablation, this is close to a clear accept at a strong venue. I do not see fabrication or circular scoring beyond standard delayed-feedback online protocols. Fit for cs.CL / multimodal agent tracks is good."},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"Punchline: this is a solid engineering paper on top of M3-Agent. The real move is treating historical trajectories as procedural retrieval control—success/failure banks, anomaly-triggered reuse, answers grounded only in current-video evidence—not as another fact dump into the prompt.\n\nWhat is new is the separation, not the ingredients. Entity graphs, verbal reflection, and lifecycle hygiene already exist. RRM’s contribution is putting them in the right place: Online Query Reflection for within-task repair, experience only as query focus after stripping answers/entities, cross-video exclusion, and delayed mini-batch feedback so experience cannot leak within a video. The evidence package is coherent: same factual backbone and answer model as M3-Agent; progressive ablations (OQR → experience → lifecycle); query-only beating prompt injection; accuracy up and retrieval rounds down on Robot, Web, and Video-MME-Long.\n\nSoft spots, in proportion. The stress-test is right that they never directly audit residual leakage in extracted records or run a poisoned-experience / cross-domain transfer check. So you cannot treat the 9.1/5.8/7.4 gains as pure proof of “transferable procedural knowledge.” A weaker reading—“better retrieval controller under anomaly triggers”—still fits the data and is still useful. Other limits are ordinary for this venue: means of three without error bars, LLM judge, free lifecycle and batch hyperparameters, no released code in the manuscript. None of that collapses the central design claim.\n\nCitations look appropriate (M3-Agent, Reflexion/ExpeL/AWM/ReMe, long-video baselines). No fake math; this is empirical systems work.\n\nWho it is for: people building long-horizon multimodal agents and memory-augmented retrieval who care about where experience enters the loop. Worth a serious referee. I would engage—read the ablations and the query-only vs injection figure carefully, and ask for a leakage audit or code in revision—not dismiss it.","headline":"Clean systems win on query-only experience reuse for long-video agents; isolation is designed, not audited, but the controlled gains still hold.","tokens_in":14176,"tokens_out":507,"would_cite":true,"duration_ms":18211,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.5","headline":"Long-video agents get better by learning how to search, not by replaying old answers.","keywords":["reflective retrieval memory","long-horizon multimodal reasoning","procedural experience","query-only reuse","entity-centric memory graph","long-video understanding","online lifecycle management"],"falsifier":"Run the same benchmarks with a leakage audit: if query-only reuse still injects source-video entities or answers into later queries, or if gains vanish when experience is drawn only from truly disjoint videos with no ground-truth-aided failure extraction, the central claim fails.","tokens_in":13962,"feed_emoji":"🧠","tokens_out":881,"duration_ms":16231,"temperature":0.7,"pith_summary":"Long-horizon multimodal agents usually stockpile facts from long videos but still fail when retrieval goes wrong and stay stuck repeating bad search habits. This paper argues the missing piece is reusable procedural knowledge about how to search: what evidence a question needs, which query patterns work, and how to correct failed retrieval. Reflective Retrieval Memory (RRM) keeps a separate experience store distilled from past successful and failed task trajectories, turns selected experiences into query-level guidance only, and still answers solely from newly retrieved current-video facts. A lifecycle manager prunes and merges experiences by reuse, feedback, and time so the store stays useful. On three long-video benchmarks the method beats strong memory agents while using fewer retrieval rounds, supporting the claim that learning how to retrieve improves both accuracy and efficiency.","feed_headline":"Agents learn how to search long videos, not what to recall","feed_subtitle":"Query-only reuse of past retrieval strategies beats strong memory baselines and needs fewer search rounds.","key_machinery":"Reflective Retrieval Memory (RRM): a third memory layer of structured successful/failure procedural records, gated by Online Query Reflection and anomaly-triggered query-only reuse, plus lifecycle management by usage, reuse feedback, and temporal decay, so experience steers queries without entering the answer context.","core_discovery":"The paper establishes that augmenting an entity-centric multimodal memory graph with reflective experience memory—procedural retrieval knowledge distilled from historical trajectories and reused only as query-level control, with answers conditioned only on newly retrieved current-video evidence—consistently improves long-horizon multimodal reasoning, raising scores over the shared-backbone baseline by 9.1, 5.8, and 7.4 points on M3-Bench-Robot, M3-Bench-Web, and Video-MME-Long while cutting average retrieval rounds.","pith_inferences":["If procedural retrieval memory generalizes, streaming robot and web agents could improve mid-deployment without enlarging the video context window.","Stricter automated leakage tests (entity overlap between experience text and current-video answers) would be a natural next measurement the paper leaves open.","The success/failure split suggests asymmetric trust policies may matter in other agent memory systems beyond video QA.","Reducing retrieval rounds at higher accuracy implies experience reuse could cut inference cost in multi-round tool-using agents."],"forward_implications":["Long-video agents should store search procedures separately from video facts and never feed historical experience into the answer generator.","Failed retrieval trajectories are useful mainly as corrective query strategies, not as full episodes to replay.","Within-task anomaly detection plus cross-task experience can raise accuracy while reducing retrieval rounds.","Experience stores need online lifecycle control (merge, decay, prune) or they accumulate noise that hurts later tasks.","The same query-only reuse pattern can be stacked on other factual memory graphs without redesigning what is stored from the current video."],"fun_headline_variants":["Reflective memory reuses search tactics, not old facts","Agents improve video search via procedural retrieval memory","Query-only past strategies lift long-horizon reasoning scores","Experience memory cuts retrieval rounds on long videos","Entity graph plus reflective search beats strong baselines"],"cache_read_input_tokens":128,"weakest_assumption_plain":"The load-bearing premise is that LLM-extracted “how to search” records, after stripping answers and entities and blocking same-video reuse, transfer as pure retrieval control without leftover task-specific leakage under the mini-batch delayed-feedback protocol.","fun_headline_variants_meta":{"raw":{"variants":["Reflective memory reuses search tactics, not old facts","Agents improve video search via procedural retrieval memory","Query-only past strategies lift long-horizon reasoning scores","Experience memory cuts retrieval rounds on long videos","Entity graph plus reflective search beats strong baselines"]},"model":"grok-4.5","effort":"low","cost_usd":0.002657,"raw_usage":{"total_tokens":1048,"prompt_tokens":799,"num_sources_used":0,"completion_tokens":55,"cost_in_usd_ticks":26568000,"prompt_tokens_details":{"text_tokens":799,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":194,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":799,"tokens_out":55,"duration_ms":4866,"temperature":1.0,"reasoning_tokens":194,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-31T16:07:27.069829+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"Run the same benchmarks with a leakage audit: if query-only reuse still injects source-video entities or answers into later queries, or if gains vanish when experience is drawn only from truly disjoint videos with no ground-truth-aided failure extraction, the central claim fails.","supporting_citations":[],"review_version":1}