{"id":"cbc93cff-2405-4e1c-a831-996364117ff3","arxiv_id":"2608.07622","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":8,"one_line_summary":"Memory interference between stored experiences, especially repeated history and same-slot conflicts, sharply suppresses update adoption in continual LLM agents, and controlled conflict examples can be used to train better memory retrieval.","lead":"This paper introduces a controlled test framework showing how conflicting memories, not just the total amount of stored information, cause LLM agents to fail at adopting new, correct updates. It also shows that adding these conflict cases to training data improves agents' ability to distinguish valid updates from misleading memories without hurting original-task performance.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Same-slot conflict memories are appended after the target, so under the 'latest applicable stable update' instruction they may constitute the correct output; the plasticity drop may be partly an instruction artifact rather than relationship-specific interference.","rationale":"The reader's weakest assumption is the preservation claim in Section 3.2: that controlled additions leave the underlying task unchanged, so drops in plasticity must be caused by the memory relationships. I agree this assumption is load-bearing, and I identify a specific failure mode: the composition protocol preserves target position by appending additions after the target, while the Travel prompt instructs the model to use the 'latest applicable stable update.' A same-slot conflict memory appended later than the target is therefore a plausible 'latest' update, especially when repeated eight times. The model's adoption of that value could be correct behavior under the prompt, not a plasticity failure. The paper's P0–P3 authority probe partly addresses this by showing recency alone has little effect, but it does not cover the multi-repetition conflict condition, and the paper does not specify whether conflict memories are marked as transient or otherwise invalid. This is a concrete, testable threat to the central claim that relationship rather than count determines the failure regime. I do not move the verdict to reject because the paper has multiple independent lines of evidence (repeated history, external growth, retrieval diagnostics) that are less affected by this confound, and the proposed ablation would settle the issue. The reader's CONDITIONAL verdict is appropriate; the insertion-order and transient-wording tests should be conditions for final acceptance.","tokens_in":23330,"tokens_out":16000,"duration_ms":151789,"concrete_test":"Run the TravelPlanner same-slot conflict ablation with insertion order reversed so the target memory is placed after the conflict memories, keeping content and count fixed. If the count-8 plasticity loss largely disappears (e.g., moves from −0.50 toward the unrelated condition's −0.02), the effect is an interaction between the prompt's 'latest applicable stable update' rule and the fixed target-position preservation, not purely relationship-specific interference. As a second check, annotate the conflict memories with transient/one-off wording (as in the mirrored stability protocol) and verify whether the plasticity drop persists under a prompt that instructs ignoring single-occasion values.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 3.2 states that all CMI variants preserve the query, expected decision, and target-memory set, and additionally preserve target position. In the relationship-controlled composition protocol (C.2), controlled additions are appended to the base collection while the target position is preserved, so same-slot conflict memories are temporally later than the target. The Travel composition prompt (D.3) instructs: 'If memories conflict, use the latest applicable stable update for the same traveler/day/slot.' A later memory that expresses a competing value for the same slot is therefore a candidate for 'latest applicable stable update.' If a model adopts that value, it is following the prompt's decision rule, not necessarily failing to be plastic. The paper's fixed-L3 authority probe (P0–P3) shows that update-authority wording, not recency alone, drives acceptance of an invalid memory; but the composition condition supplies up to eight conflicting repetitions, which may collectively be read as a stable update. The assertion that y_i is preserved assumes the conflicting value is not a valid update under the prompt's rule, yet the paper does not report whether the conflict memories carry transient/exception markers or are otherwise distinguishable from valid updates. Without this, the count-8 same-slot plasticity drops (−0.50 lexical, −0.44 dense) could reflect the model rationally following 'latest' rather than memory interference.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces Controlled Memory Interference (CMI), a diagnostic and data-generation framework for studying how relationships among accumulated memories affect continual LLM agents. CMI varies memory load and relationship type while preserving the query, target state, and expected decision, and evaluates plasticity, stability, and poisoning resistance. The authors report that benign accumulation has limited, non-monotonic effects, whereas repeated history and same-slot conflict sharply suppress plasticity, with lexical and dense retrieval exhibiting distinct interference pathways. They also show that CMI-generated training examples improve interference-aware retrieval and answer success across five datasets, and they provide extensive reproducibility artifacts including audits, bootstrap intervals, and explicit limitations.","tokens_in":23574,"tokens_out":8445,"duration_ms":81652,"significance":"If the central claim holds, this is a valuable and well-engineered contribution to the study of memory in continual LLM agents. The paper's strengths include the paired base-cluster bootstrap, the retrieval-to-decision decomposition, the multi-encoder robustness check, the careful documentation of data-construction audits, and the honest statement of evidence boundaries. The framework is portable and the augmentation results are practically relevant. However, the central claim that memory relationships, rather than count, determine the failure regime is threatened by a prompt-level confound that must be resolved before the conclusion can be accepted.","major_comments":[{"comment":"The relationship-controlled composition protocol appends the controlled additions after the target memory while preserving the target position, and the generation prompt instructs the model to 'use the latest applicable stable update for the same traveler/day/slot.' A same-slot conflict memory that appears later in time and expresses a competing value is therefore a candidate for 'latest applicable stable update,' and the paper does not report whether these conflict memories carry transient, one-off, or otherwise invalid markers. The fixed-L3 authority probe (P0–P3, Table J.5) shows that recency alone is weak, but that probe uses a single invalid memory; the composition sweep supplies up to eight repetitions, which may collectively convey stability. Consequently, the measured plasticity drops (Lexical −0.50 for same-slot conflict and −0.34 for repeated history at count 8; Dense −0.44 and −0.46) and the P(A|T)<1 values in Table 2 could partly reflect the model rationally following the prompt's 'latest' rule rather than memory interference. The central claim that 'relationship, rather than count, determines the failure regime' requires a control in which the appended conflict/repetition memories are explicitly marked as transient or invalid, or a prompt that disambiguates update status.","section":"Section 3.2 and Section D.3"},{"comment":"The external L0–L6 growth protocol accumulates same-slot historical states, and the answer-generation prompt only says 'Resolve persistent updates and one-off exceptions carefully.' If the generated historical states are phrased as plain assertions without explicit 'superseded' or 'temporary' markers, a model that retrieves a later old-state memory may treat it as the current persistent update. Under this ambiguity, the large lexical plasticity drops (0.72–0.94 on PERMA, LongMemEval, HorizonBench, and MEME) could reflect the interaction between retrieval exposure and the model's update inference rather than relationship-specific interference alone, while the smaller dense drops would then be an artifact of which memories are retrieved. The authors should report whether the historical-state texts contain update language or markers, and ideally add a control condition that explicitly frames the accumulating memories as historical/superseded.","section":"Section 5.3, Table 3, and Appendix K"},{"comment":"The paper states that 'the standalone PERMA and LongMemEval answer summaries were recovered from an audited consolidated evidence record' and that some intervals and absolute wrong-use values were not retained. Since Table 3 and the cross-dataset convergence argument rely on these answer-level summaries, the provenance of the recovered numbers should be made fully transparent: which CSV rows are original and which were recovered, and why the per-sample outputs for those sources are missing. Without this, the audit trail for the cross-dataset convergence claim is weaker than for the TravelPlanner results.","section":"Appendix L"}],"minor_comments":[{"comment":"Section 4.1 states that TravelPlanner has '1,000 fixed-L3 poisoning instances,' but Section C.3 and Table J.5 describe 100 paired cases per access mechanism per condition, which totals 800 (2 mechanisms × 4 conditions × 100). The discrepancy should be reconciled.","section":"Section 4.1 vs. Section C.3"},{"comment":"The term 'architecture-dependent' is used to describe differences between lexical and dense retrieval. Since the underlying LLM is fixed, this is better described as 'access-mechanism-dependent' or 'retrieval-family-dependent' to avoid implying a difference in model architecture.","section":"Throughout"},{"comment":"The reference to Voorhees and Tice appears as 'V oorhees' in the references list; please fix the spacing.","section":"References"},{"comment":"The encoder-robustness section would benefit from a short paragraph interpreting the non-monotonic count effects (e.g., NV-Embed showing zero loss for matched neutral at all counts) rather than leaving the reader to extract the pattern from the figure alone.","section":"Appendix J.10 / Figure J.9"}],"recommendation":"major_revision","confidential_remarks":"This is a careful and unusually transparent paper, and the CMI framework is likely to be useful to the community. The main obstacle is the prompt-level confound in the relationship-controlled composition and external growth protocols: without explicit markers of invalid/transient status, the 'latest applicable stable update' instruction may itself produce the observed plasticity suppression. This is addressable with a control condition or a modified prompt, so I recommend major revision rather than rejection. The recovered-summary transparency issue for PERMA and LongMemEval should also be resolved. If the authors can close the prompt confound, the paper would be a strong accept."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"CMI is a useful, carefully built framework for studying memory interference in continual agents, and the cross-dataset augmentation results are honestly scoped. But the headline composition result has a confound that needs addressing: same-slot conflict memories are appended after the target, and the prompt tells the model to use the latest applicable stable update. Under that instruction, a later memory expressing a competing value is a valid answer, so the observed plasticity drop is not cleanly attributable to interference.\n\nWhat is genuinely new here: the paired relationship-controlled scenarios, the retrieval-to-decision decomposition, and the observation that repeated history suppresses update plasticity even when it introduces no new value. The experimental design is above average for this subfield—cluster-bootstrapped paired intervals over base scenarios, audited transformations, separate retrieval and answer protocols, and a transparent limitations section. The augmentation claim is modest and in-domain, and the paper explicitly avoids overclaiming transfer.\n\nThe soft spots are proportionate. The same-slot confound is real: Appendix D.3's instruction says to use the latest applicable stable update, and the paper's own P0–P3 probe shows that update authority, not recency, drives acceptance. Eight repetitions of a competing value can easily be read as a stable update. The repeated-history condition, which only restates the prior value, shows a smaller but still substantial plasticity drop, so the broader phenomenon is probably real; but the same-slot magnitudes in Table 1 are likely inflated. The stability protocol is cleaner because it marks conflicts as transient, and the dense stability drop there is genuine interference, but that does not rescue the plasticity interpretation.\n\nOther weaknesses are minor and disclosed: some external summaries were recovered from an audited evidence record, the API generator does not expose a sampling seed, and a few intervals are unavailable. None of these are load-bearing.\n\nI would send this to a serious referee. The framework is novel, the experiments are reproducible enough to be checked, and the confound is fixable: either place conflicts before the target, or mark them explicitly as transient, or add a control that separately varies update authority. With that, the CMI diagnostic would be a solid contribution to agent-memory research.","headline":"Novel, careful CMI framework for agent-memory interference; main same-slot-conflict effect is partly confounded by the prompt's latest-update rule.","tokens_in":24119,"tokens_out":6359,"would_cite":true,"duration_ms":55248,"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":"In continual LLM agents, memory interference is relationship-specific: repeated history and same-slot conflicts suppress valid update adoption even though memory scale alone barely matters, and controlled conflict cases can be used to…","keywords":["memory interference","continual LLM agents","selective plasticity suppression","long-term memory","retrieval-to-decision decomposition","memory poisoning","interference-aware retrieval","controlled memory evolution"],"falsifier":"Force the target memory into the retrieved context at a fixed position regardless of rank, under the lexical same-slot-conflict condition at count 8: the paper's decomposition predicts answer success recovers to roughly the 0.84 baseline because $\\Pr(A \\mid T) = 1.00$ there, so if adoption stays depressed the interference also acts downstream and the before/after split is wrong. As a cross-check, if eight token-overlap-matched but slot-unrelated additions reproduced the same-slot drop of 0.50 in plasticity, the relationship account would reduce to surface-overlap pressure.","tokens_in":23095,"feed_emoji":"🧠","tokens_out":12709,"duration_ms":98706,"temperature":0.7,"pith_summary":"This paper tries to establish that failures in continual LLM agents to adopt valid memory updates are caused by the relationships among accumulated memories — repeated history and same-slot conflict — rather than by the sheer number of stored memories. It introduces Controlled Memory Interference (CMI), a diagnostic protocol that adds memories of different relationship types while holding the query, target memory, and expected answer fixed, and finds that benign accumulation barely changes behavior while relationship-specific interference sharply suppresses update plasticity (lexical plasticity drops of 0.34 under repeated history and 0.50 under same-slot conflict at eight additions) with little stability gain. The paper further shows that the effect is retrieval-architecture dependent, can occur either before or after the target memory reaches the model's context, and that CMI-generated training cases improve noisy-update performance across five datasets with no observed decrease on matched clean estimates. A sympathetic reader would care because the result reframes continual-agent memory failures as an interaction problem rather than a capacity problem, and offers a portable way to diagnose and train against the failure.","feed_headline":"Memory conflicts, not size, break LLM agent updates","feed_subtitle":"Repeated or same-slot memories cut update adoption by up to 0.50; neutral accumulation barely matters.","key_machinery":"The load-bearing mechanism is CMI itself: a paired-scenario construction that fixes the query, target memory, target position, and expected decision while varying memory load (L0–L6) and relationship composition (matched neutral, unrelated, repeated history, same-slot conflict), plus a fixed-L3 authority probe that frames an invalid memory as noise, a recent mention, a persistent update, or an authoritative correction. Around it, the paper defines a retrieval-to-decision decomposition — target exposure $T$, competitor exposure $C$, and conditional answer success $\\Pr(A \\mid T)$ and $\\Pr(A \\mid T, C)$ — that tells whether interference blocked access or disrupted downstream use. For the training use, an Interference-Aware Adaptive Retrieval (IAAR) selector, a random forest over observable retrieval-state features (score margins, retrieval entropy, lexical overlap, duplicate density, cross-retriever agreement), predicts a mixture weight $\\alpha$ between lexical and dense scores, trained either on clean data or augmented with CMI noisy same-slot cases.","core_discovery":"The central discovery is selective plasticity suppression: within fixed memory counts, the relationship between an added memory and the target state determines when an agent fails to adopt a valid update, with 'relationship, rather than count, determining the failure regime.' Matched-neutral and unrelated additions leave plasticity nearly intact, while repeated history and same-slot conflict cut lexical plasticity by 0.34 and 0.50 and dense plasticity by 0.46 and 0.44 at count 8; stability stays essentially flat except for dense same-slot conflict, which drops stability from 0.92 to 0.54. Because noise-only growth from L0 to L6 moves plasticity only from 0.510 to 0.455, the suppression is not a capacity or scale effect, and because stability does not mirror plasticity, it is not a symmetric stability–plasticity trade-off. A retrieval-to-decision decomposition localizes the failure: lexical same-slot conflict is purely a retrieval bottleneck (target recall 0.34, but $\\Pr(A \\mid T) = 1.00$), lexical repeated history degrades both access and adoption, and dense retrieval fails at both stages. The same selective suppression appears across PERMA, LongMemEval, HorizonBench, and MEME Tracking, and in-domain CMI augmentation lifts noisy-update answer success by 0.114 to 0.614 across all five sources.","pith_inferences":["If the relationship-specific account is right, the remedy for update-adoption failures is conflict-aware ranking — detecting that a retrieved memory occupies the same entity–attribute slot as the target — rather than bigger or better-summarized memory stores; the paper demonstrates the detection in-domain but stops short of prescribing it as a general architecture rule.","The lexical same-slot result implies many update failures are retrieval-rank problems rather than reasoning failures; the prediction that oracle-exposure injection restores baseline performance is a cheap follow-up experiment the paper does not run.","The sensitivity of poisoning to 'persistent update' and 'authoritative correction' phrasing suggests the fixed-L3 protocol could double as a robustness audit for prompt-injection and instruction-override defenses, since it isolates surface authority cues from genuine recency.","Because the frozen selector does not transfer across embedding backends while retrained selectors recover, the semantics of the observable retrieval features appear backend-specific; a natural extension is training on a mixture of backend features to obtain a transferable policy."],"forward_implications":["Memory scale alone does not explain update-adoption failures: noise-only growth from L0 to L6 shifts plasticity only from 0.510 to 0.455, while eight relationship-specific additions can cut it by about half; continual memory systems should therefore track relationship composition, not just volume.","For lexical same-slot conflict the entire failure is target non-exposure ($\\Pr(A \\mid T) = 1.00$), so ranking interventions that guarantee target inclusion are a sufficient fix for that failure mode.","Because conveyed update authority, not recency, drives poisoning acceptance (0.23–0.26 under update wording versus 0.02–0.00 under recency-only), recency-based defenses are insufficient; agents must parse the update role a memory claims.","In-domain CMI augmentation improves target recall by 0.114–0.614 and noisy-update answer success by 0.114–0.614 across all five sources, with no observed decrease on matched clean point estimates, so controlled conflict cases are usable training data.","The qualitative interference pattern persists across four dense encoders in direction, if not magnitude, so the relationship effect is not tied to one embedding model."],"supporting_citations":[{"why":"Supplies the MemoryArena TravelPlanner source cases that CMI transforms into controlled memory trajectories.","marker":"He et al., 2026"},{"why":"Defines BM25, the lexical retrieval family whose sensitivity to repeated surface cues and same-slot conflict carries the architecture-dependence claim.","marker":"Robertson and Zaragoza, 2009"},{"why":"Provides Qwen3-Embedding-8B, the primary dense encoder whose semantic-crowding pathway is compared against lexical retrieval.","marker":"Zhang et al., 2025"},{"why":"Supplies the proactive-interference concept the paper frames memory interference with, motivating relationship-based rather than count-based failure.","marker":"Underwood, 1957"},{"why":"Supplies the cue-overload mechanism used to explain why shared retrieval cues among related memories become less discriminative.","marker":"Watkins and Watkins, 1975"},{"why":"Contributes the PERMA dataset used in the cross-dataset convergence test.","marker":"Liu et al., 2026"},{"why":"Contributes LongMemEval, one of the four external sources confirming selective plasticity suppression.","marker":"Wu et al., 2025"},{"why":"Contributes HorizonBench, an external source for the same-slot historical growth protocol.","marker":"Li et al., 2026"},{"why":"Contributes the MEME Tracking task, the fourth external source in the convergence analysis.","marker":"Jung et al., 2026"}],"fun_headline_variants":["Memory type, not volume, blocks LLM agent updates","Same-slot memories stall LLM update plasticity","Selective memory interference, not scale, impairs LLM agents","Relationship, not count, dictates LLM update failure"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The claim rests on the assumption that adding controlled memories (repeated histories or same-slot conflicts, for example) leaves the task itself unchanged, so that every measured drop in update adoption must come from the memory relationships rather than from confounds such as longer context, altered lexical overlap, or harder retrieval; the paper states this preservation premise in Section 3.2, and if it fails the relationship-specific effects could be artifacts of changed task difficulty.","fun_headline_variants_meta":{"raw":{"variants":["Memory type, not volume, blocks LLM agent updates","Same-slot memories stall LLM update plasticity","Selective memory interference, not scale, impairs LLM agents","Relationship, not count, dictates LLM update failure"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000314,"raw_usage":{"total_tokens":1830,"prompt_tokens":1039,"completion_tokens":791,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":655,"completion_tokens_details":{"reasoning_tokens":725}},"tokens_in":655,"tokens_out":791,"duration_ms":6842,"temperature":1.0,"reasoning_tokens":725,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T00:28:28.748435+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Force the target memory into the retrieved context at a fixed position regardless of rank, under the lexical same-slot-conflict condition at count 8: the paper's decomposition predicts answer success recovers to roughly the 0.84 baseline because $\\Pr(A \\mid T) = 1.00$ there, so if adoption stays depressed the interference also acts downstream and the before/after split is wrong. As a cross-check, if eight token-overlap-matched but slot-unrelated additions reproduced the same-slot drop of 0.50 in plasticity, the relationship account would reduce to surface-overlap pressure.","supporting_citations":[{"cited_title":"The Probabilistic Relevance Framework:","cited_arxiv_id":null,"evidence_quote":"Defines BM25, the lexical retrieval family whose sensitivity to repeated surface cues and same-slot conflict carries the architecture-dependence claim."}],"review_version":1}