{"id":"2c58d501-7d27-4f34-90ec-089a5a16cb3e","arxiv_id":"2607.29440","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"ADAMM induces queryable analytic tables from multimodal interaction histories and combines them with semantic retrieval, improving benchmark accuracy by up to 11.3 points over memory baselines.","lead":"This paper introduces ADAMM, a memory framework that lets AI agents both retrieve past multimodal interactions and compute over them—answering questions like \"average sleep this month\"—by inducing queryable tables from observations. On two long-term memory benchmarks it reports gains over retrieval-only baselines, with the clearest wins on analytical questions.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Extractor accuracy in §3.3.1 is the load-bearing assumption: without measurement or sensitivity analysis, the reported gains could be artifacts of extraction errors or unreported thresholds.","rationale":"The paper's central claim is that analytic memory improves multimodal QA. The mechanism is plausible and the benchmark improvements are consistent, with the ablation showing that removing analytic memory drops Health accuracy by 14.9%, directly tying the gain to the analytic components. However, the entire analytic memory rests on the LLM Extractor's output. Given that extraction from multimodal content is error-prone, and the paper provides no measurement of extraction quality or sensitivity analysis, the reported gains could be sensitive to extraction noise. The paper itself concedes this. This is not a fatal flaw but a condition on the claim: the claim holds only if the extractor is sufficiently accurate and the thresholds are reasonable. Since thresholds are unreported, reproducibility is also in question. I agree with the reader's identification of the weakest assumption. The concrete test would quantify extraction error and its impact, which would resolve whether the concern lands. Therefore, the verdict remains CONDITIONAL, and no adjustment is needed.","tokens_in":11058,"tokens_out":6972,"duration_ms":74266,"concrete_test":"Annotate 100 interaction rounds from MemEye with gold attribute-value pairs; run the Extractor and compute precision/recall/F1. Then re-run the MemEye evaluation after randomly corrupting or deleting extracted tuples at the measured error rate. If the overall LLM-Judge accuracy drops by more than 2 absolute points, or if extraction F1 is below 90%, the central claim is not robust to realistic extraction noise.","verdict_should_be":"UNCHANGED","load_bearing_attack":"ADAMM's analytic memory pipeline is built on the Extractor's output (Eq. 1). The Extractor is an LLM that must 'identify every grounded attribute–value correspondence' from dialogue and images. If it misses attributes (low recall) or fabricates values (low precision), the support/all-confidence/extension-confidence statistics (Eqs. 3–7) induce incorrect schemas, tables contain wrong or blank cells, and COMPUTE results are biased. The paper acknowledges this in Limitations ('Incorrect or missing fields may propagate...') but reports no extraction accuracy, no error analysis, and no sensitivity study. Moreover, the thresholds θ_s, θ_a, θ_e and weight λ are never reported, so the results cannot be reproduced or checked for threshold sensitivity. Without evidence that extraction errors are rare and do not materially affect downstream answers, the reported gains (up to 11.3%) could be an artifact of the specific extraction quality or threshold choices rather than a robust property of analytic memory.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper argues that existing multimodal agent memory systems, organized around relevance-based retrieval, are poorly suited to analytical questions that require complete records and operations such as filtering, aggregation, ranking, and temporal comparison. It introduces ADAMM, which adds an \"analytic memory\" layer to standard retrieval memory. The construction pipeline has three stages: (i) an LLM-based extractor identifies grounded attribute--value pairs from each interaction round (Eq. 1); (ii) a schema-induction procedure mines recurring attribute co-occurrences using support, all-confidence, and extension-confidence (Eqs. 3--7); and (iii) each induced schema is materialized as a relational table, while retrieval memory is organized as a topic--episode--event hierarchy. At inference, a planner first builds a query-conditioned metadata context (Eq. 9), generates a high-level plan, and then progressively instantiates tool calls whose later arguments depend on earlier outputs. The paper evaluates ADAMM on MemEye and MemGallery with two backbone models, reporting consistent improvements over unimodal, multimodal, and RAG baselines, with headline gains of up to 11.3 and 7.3 percentage points. An ablation on MemEye shows drops when either analytic or retrieval memory is removed and when planning context or progressive execution is ablated.","tokens_in":11270,"tokens_out":3112,"duration_ms":38733,"significance":"If the empirical claims hold, the paper makes a useful conceptual contribution by separating analytic memory from retrieval memory and by showing that schema induction over LLM-extracted observations can support executable analysis in multimodal settings. The framework is clearly motivated, the schema-induction formalism follows standard frequent-itemset and all-confidence ideas, and the evaluation is broader than many memory papers: two benchmarks, six metrics, two backbones, and ablations. The fine-grained task analysis in Figure 3 gives a plausible mechanistic account of where the gains come from (tasks requiring exact operations over recurring records). The paper also transparently acknowledges its main vulnerability: extraction errors propagate through schema induction and downstream computation. However, this limitation is never quantified, and the manuscript does not report key thresholds or run-level variance. These omissions make it difficult to assess whether the reported gains are robust properties of analytic memory or artifacts of particular extraction quality, threshold choices, or evaluation noise. The central idea is defensible and worth publishing, but the current","major_comments":[{"comment":"The pipeline's load-bearing assumption is that the LLM Extractor identifies 'every grounded attribute–value correspondence' with sufficient precision and recall. The manuscript reports no extraction accuracy, no error analysis, and no sensitivity study on how imperfect extraction affects schema induction, table population, or downstream COMPUTE results. The Limitations section itself concedes that 'Incorrect or missing fields may propagate to schema induction, table construction, and downstream computations.' Since the paper's central claim is that analytic memory produces the observed gains, the authors should provide at least a manual or automatic evaluation of extractor precision/recall, and ideally a perturbation analysis (e.g., adding/deleting extracted fields) to show the results are not driven by extraction artifacts.","section":"§3.3.1, Eq. (1); Limitations"},{"comment":"The main results are averaged over only 3 runs, and no standard deviations, confidence intervals, or significance tests are reported. Several margins are small (e.g., 2.2–2.6 percentage points for MemGallery with GPT-5.4-mini), so without run-level variability it is impossible to know whether these differences are reliable. The paper repeatedly claims 'consistent' improvements; this requires statistical support. Please report per-run results or error bars and perform significance tests where appropriate.","section":"§4.1, Table 3"},{"comment":"The schema-induction thresholds θ_s, θ_a, θ_e and the metadata-ranking weight λ are never reported. These parameters control which schemas are discovered, whether schemas evolve, and which planning candidates are exposed to the planner. Without their values, the experiments cannot be reproduced, and the reader cannot judge whether the results are sensitive to particular threshold choices. The authors should report the exact thresholds used, and ideally perform a sensitivity analysis over reasonable ranges.","section":"§3.3.2, Eqs. (4), (6), (7); §3.5.2, Eq. (9)"},{"comment":"The ablation supporting the complementarity of analytic and retrieval memory is conducted only on MemEye, and only two task categories (Health and Brand) are shown. The paper's claims about general benefit across benchmarks would be stronger if the same ablation were run on MemGallery, especially on tasks such as Conflict Detection and Knowledge Resolution where structured access is said to matter most. As it stands, the complementarity claim is supported on only one benchmark.","section":"§4.3, Figure 4"}],"minor_comments":[{"comment":"The paper does not provide the exact prompts used for the Extractor, planner, metadata generation, or answer model. These prompts are essential for reproducibility, particularly since the pipeline is LLM-driven.","section":"General"},{"comment":"The category abbreviations in Figure 3 (AR, CD, FR, KR, MRTR, TTL, VR, VS) are not defined in the caption or text. They should be expanded or listed in a legend.","section":"Figure 3"},{"comment":"Typos and formatting issues: 'systems emphasizes' in the Introduction; 'MemGalleryreports' in §4.1; 'consistant' in §4.2; 'retrieval memory' tool row in Table 1 is missing a capital letter. These do not affect substance but should be corrected.","section":"General"},{"comment":"The sentence 'we use Top-10 retrieved memory units for retrieval memory baselines' is ambiguous: is this a budget for all baselines, and if so, how does it interact with ADAMM's shared budget of 10 evidence units? Clarify the comparison protocol.","section":"§4.1"}],"recommendation":"major_revision","confidential_remarks":"The paper is well within the scope of the journal and the core idea is sound, but the empirical support is currently incomplete in ways that directly affect the headline claim. The missing extraction-accuracy analysis is particularly important because the authors themselves identify it as the main risk. I would encourage the editor to request a revision that adds error bars/statistical tests, reports the missing thresholds, and provides some evidence about extractor reliability rather than a simple restatement of the limitation."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Punchline: This paper has a genuine new abstraction and a clean empirical story, but the load-bearing component—LLM extraction of attribute-value pairs—is never measured, and key thresholds are missing. Worth refereeing, not desk-rejecting.\n\nWhat's new: ADAMM's analytic memory is not just another retrieval store. It extracts provenance-linked attribute-value observations, induces recurring schemas via support/all-confidence/extension-confidence, materializes tables, and plans queries over both retrieval and analytic tools. That's a real departure from the retrieve-then-answer baselines. The formulation of the retrieval-analysis mismatch is useful and likely to be reused.\n\nWhat it does well: The experiments are broad (two benchmarks, two backbones, six metrics, ablations) and the improvement pattern makes sense—largest gains on tasks requiring aggregation and comparison (Health, Card Playlog, Conflict Detection), smaller gains on pure visual recall. The ablations show both memory components contribute, and progressive execution helps when later tool calls depend on earlier results. The paper is clearly written and honest in its limitations.\n\nSoft spots: The extractor in §3.3.1 is the foundation. If it misses attributes or fabricates values, the schema statistics (Eqs. 3–7) and tables inherit those errors. The paper acknowledges this but reports no extraction accuracy, no error analysis, no sensitivity to thresholds. And the thresholds θ_s, θ_a, θ_e and λ are simply not given, so the pipeline cannot be reproduced or checked for threshold sensitivity. No error bars or significance tests either, though the consistent direction across metrics and backbones makes the qualitative claim credible. Minor: the abstract says \"up to 7.3%\" on MemGallery, but Table 3's largest MemGallery gain is 7.0; a small copy-paste slip, but worth fixing. No code, prompts, or data artifacts are released, which matters for a pipeline whose core component is prompt-driven.\n\nBottom line: The architecture is plausible, the central claim—analytic memory complements retrieval—is supported by the evidence as reported, and the limitations are addressable rather than fatal. This is a paper for anyone working on long-term agent memory, especially multimodal settings. It deserves peer review; the referee should ask for extraction-quality measurements, threshold reporting, and ideally code/data release.","headline":"ADAMM makes a real architectural claim—complementing retrieval with induced analytic tables—and backs it with consistent gains, but the unmeasured extractor and unreported thresholds keep it from being fully convincing yet.","tokens_in":11765,"tokens_out":2172,"would_cite":true,"duration_ms":22848,"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":"A memory that extracts recurring attribute-value pairs from dialogues and images, induces schemas, and materializes them into queryable tables lets agents answer analytical questions that pure retrieval cannot.","keywords":["multimodal memory","analytic memory","retrieval-analysis mismatch","schema induction","attribute-value extraction","memory-aware planning","agent memory","long-term interactions"],"falsifier":"Measure extraction precision and recall on a human-annotated subset of the Health task, then corrupt or delete a random fraction of extracted fields and check whether table-based COMPUTE answers degrade accordingly. If deletion barely affects accuracy, the analytic-memory advantage would not be attributable to correct structure; if accuracy degrades sharply, the extraction assumption is load-bearing.","tokens_in":10927,"feed_emoji":"📊","tokens_out":4123,"duration_ms":43499,"temperature":0.7,"pith_summary":"The paper argues that long-term multimodal memory has a retrieval-analysis mismatch: retrieval-based systems return relevant snippets but cannot compute over complete, correctly scoped records. It introduces analytic memory as a complement, built by extracting attribute-value observations from dialogue and images, inducing recurring schemas via support and confidence thresholds, and materializing them as tables. At query time, a planner selects and composes retrieval and analytic tools such as lookup, filter, compute, and rank. On two long-term multimodal memory benchmarks and two backbone models, the combined system outperforms retrieval-only baselines by up to 11.3% and 7.3%, respectively. A sympathetic reader would take the core claim to be that organizing observations into induced structures is what enables analytical questions.","feed_headline":"Adding analytic tables lifts multimodal agents by up to 11.3%","feed_subtitle":"Retrieval-only memory can't average or filter the past; schema-induced analytic memory can.","key_machinery":"The carrying mechanism is a three-stage analytic memory pipeline. (1) Attribute Extraction: an LLM reads each round's dialogue and images and emits attribute-value pairs with provenance links to supporting spans or regions. (2) Schema Induction: frequent-itemset-style support counting finds recurring attribute sets; all-confidence, defined as the support of a set divided by the maximum support of its members, gates new-schema discovery; extension-confidence, defined as the support of a candidate divided by the support of the existing schema, gates schema evolution. (3) Memory Materialization: each admitted schema becomes a table whose columns are the schema's attributes plus order and time,","core_discovery":"The central claim is that relevance-based retrieval alone is insufficient for analytical queries over accumulated multimodal observation logs, and the missing capability can be supplied by schema-induced analytic memory. ADAMM first uses an LLM extractor to turn each interaction round into provenance-linked attribute-value pairs, without assuming a schema. It then mines candidate attribute co-occurrences using support, admits new schemas when all-confidence is high, extends existing schemas when extension-confidence is high, and materializes each schema as a relational table with order and time columns, leaving missing values blank. At inference, a memory-aware planner first sees metadata ab","pith_inferences":["The same schema-induction approach could be applied to other append-only multimodal logs, such as embodied navigation or tool-use traces, where recurring observations hide analytical patterns.","Because the planner only sees table metadata, not full contents, the design suggests a natural scaling path for long horizons: analytic tables grow in rows while planning context stays bounded.","A testable extension would be confidence-aware extraction: if the extractor emits per-attribute confidence, schema induction could down-weight uncertain fields before thresholds are applied, addressing the paper's stated limitation.","The retrieval-analysis distinction may generalize beyond memory: any system that both recalls and computes over evidence could benefit from explicitly separating relevance-based access from deterministic analytical access."],"forward_implications":["Analytical queries that previously required complete retrieval, such as averages, sums, rankings, and temporal windows, become answerable by deterministic table operations, reducing the coverage-context trade-off.","No application-defined schema is needed at build time; schemas emerge from recurring observations and can evolve as new attributes appear.","Queries can be decomposed into mixed retrieval-and-compute plans, so the same memory supports both semantic recall and structured analysis.","The approach generalizes across answer backbones and benchmark styles, suggesting the gain comes from memory organization rather than from a single model.","The progressive planner allows later operations to consume intermediate outputs, such as using a located event's timestamp as a constraint for a subsequent computation."],"fun_headline_variants":["Analytic memory lifts multimodal agents by up to 11.3%","Beyond retrieval: Analytic memory improves agents by 11.3%","Schema-induced analytic memory boosts multimodal agents by 11.3%","AdaMM: Analytic memory for multimodal agents, up to 11.3%","From retrieval to analytics: multimodal agents gain 11.3%"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The whole pipeline rests on the assumption that the extractor reliably finds every grounded attribute-value pair in each round; if fields are missed or wrong, the induced schemas, tables, and computed answers inherit those errors.","fun_headline_variants_meta":{"raw":{"variants":["Analytic memory lifts multimodal agents by up to 11.3%","Beyond retrieval: Analytic memory improves agents by 11.3%","Schema-induced analytic memory boosts multimodal agents by 11.3%","AdaMM: Analytic memory for multimodal agents, up to 11.3%","From retrieval to analytics: multimodal agents gain 11.3%"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00064,"raw_usage":{"total_tokens":2755,"prompt_tokens":690,"completion_tokens":2065,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":434,"completion_tokens_details":{"reasoning_tokens":1980}},"tokens_in":434,"tokens_out":2065,"duration_ms":14696,"temperature":1.0,"reasoning_tokens":1980,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-03T07:01:32.670247+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Measure extraction precision and recall on a human-annotated subset of the Health task, then corrupt or delete a random fraction of extracted fields and check whether table-based COMPUTE answers degrade accordingly. If deletion barely affects accuracy, the analytic-memory advantage would not be attributable to correct structure; if accuracy degrades sharply, the extraction assumption is load-bearing.","supporting_citations":[],"review_version":1}