{"id":"dbcd0033-0943-4b20-ba24-8b633da0859e","arxiv_id":"2608.01739","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":10,"one_line_summary":"CoEvo-Mem co-adapts a residual retrieval router and a relational memory graph with alternating phases and reports improved LLM agent performance, but it is not the top method on every benchmark in its own results.","lead":"CoEvo-Mem lets an LLM agent improve its own memory by jointly learning which past memories to fetch and how to update their value and links, alternating between the two tasks. It reports consistent gains on several agent benchmarks, though the boldest claim of state-of-the-art everywhere is undercut by its own tables.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Learned memory utility Q is never isolated by an ablation; its self-referential bootstrap (Eqs. 6-7) leaves the co-evolution benefit possibly attributable to graph/temporal structure rather than to value-based ranking.","rationale":"The reader's weakest_assumption identifies the same core issue I see: the learned Q is the only component that directly couples memory values to retrieval ranking, and it is neither isolated by an ablation nor given a stability argument. I agree with that selection over the alternative SOTA overclaim, because the false 'SOTA' statement (Table 4 shows second place on LoCoMo-MC and LongMemEval-MC) is a presentation error that can be fixed by rewording, whereas an unsupported Q-ranking mechanism would undermine the paper's central contribution. The λ_Q=0 ablation is the cleanest single experiment to settle it. The reader's CONDITIONAL verdict is appropriate: the framework is clearly specified, the alternating schedule is supported by the schedule ablation, and the failure semantics are carefully defined, but the key mechanism lacks direct evidence. I would keep the verdict CONDITIONAL and ask for the λ_Q=0 ablation plus a matched LoCoMo OpenQA comparison before elevating to ACCEPT.","tokens_in":21734,"tokens_out":14305,"duration_ms":124257,"concrete_test":"On GPQA Diamond and BFCL non-thinking, rerun the full CoEvo-Mem configuration with λ_Q set to 0 in Eq. (5), keeping SR-QR rewrites, typed relations, temporal expansion, metadata repair, and the alternating schedule exactly as in the reported runs. Compare mean and standard deviation over the same number of repeated runs. If the score with λ_Q=0 is within one standard deviation of the full model, the learned utility term is not load-bearing and the paper should de-emphasize value-based ranking; if it drops materially, the self-referential bootstrap is confirmed as a source of the benefit and needs a stability analysis.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's closed-loop claim rests on the learned utility Q changing retrieval rankings through the λ_Q/(η+rank_Q) term in Eq. (5). Q is the least isolated component: the 'w/o Memory Evolution' ablation removes typed relations, temporal expansion, and graph-based value propagation together, so its 4.80-point drop cannot be assigned to the utility signal. No experiment sets λ_Q=0 while retaining the graph, so the specific claim that learned memory values improve ranking is untested. This matters because the Q update is self-referential: Eq. (6) initializes a new memory's value to the mean of the exposed set's Q values, and Eq. (7) uses γ times that mean as the bootstrap target for the same exposed memories. Without a convergence or stability analysis, a miscalibrated Q could rank relevant memories below irrelevant ones; the system might still appear to work if dense/sparse RRF dominates (λ_Q=0.15). Thus the evidential link between memory evolution and the reported gains is the weakest step in the argument.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes CoEvo-Mem, a closed-loop framework that co-evolves a retrieval policy and a persistent memory bank for LLM agents. A frozen LLM produces route-specific query rewrites and a routing prior, and a lightweight residual router is trained online to adjust dense/sparse fusion weights. The memory side maintains a typed relational graph with learned utility values Q, updated by TD-style residuals and relational credit propagation. The two sides are trained in alternating phases, with one component fixed while the other updates. The paper evaluates CoEvo-Mem on seven benchmarks and reports large gains over baselines, with ablations showing degradation when either the query-rewriting/routing module or the memory-evolution module is removed.","tokens_in":21993,"tokens_out":5197,"duration_ms":48518,"significance":"If the central claims held as stated, the paper would make a useful empirical contribution: it identifies a feedback loop between retrieval and memory evolution and provides a concrete instantiation with careful experimental controls, including fixed splits, frozen evaluation, conservative failure semantics, and a detailed appendix. The ablation evidence that co-adaptation helps is nontrivial and generally well designed. However, the headline 'state-of-the-art across seven benchmarks' is inconsistent with the paper's own Table 4, and the specific mechanism attributed to learned memory utility is not isolated by any ablation. These issues do not invalidate the overall approach, but they currently prevent the paper from supporting the strength of its claims.","major_comments":[{"comment":"The abstract's claim that 'CoEvo-Mem achieves state-of-the-art performance' across seven benchmarks is contradicted by Table 4, where CoEvo-Mem ranks second to ElasticMem in every matched-backbone setting (e.g., 71 vs. 84 on LoCoMo-MC and 64 vs. 76 on LongMemEval-MC with Qwen2.5-7B-Instruct). The body text itself states that 'CoEvo-Mem ranks second in all matched-backbone comparisons.' The headline claim should be revised to specify the settings in which CoEvo-Mem is best, or the MC results should be explicitly reconciled with the SOTA statement.","section":"Abstract; Section 'Long-Term Conversational Memory Task'; Table 4"},{"comment":"The contribution of the learned memory utility Q to retrieval is never isolated. The 'w/o Memory Evolution' ablation removes typed relations, temporal expansion, and graph-based value propagation together, so the reported 4.80-point average drop cannot be assigned to the Q-weighted ranking term in Eq. (5). To support the specific claim that learned utilities improve retrieval rankings, the authors should add an ablation that sets lambda_Q = 0 while retaining the graph structure and propagation, or that varies lambda_Q; otherwise the gains from memory evolution could be driven entirely by temporal graph expansion or relational propagation rather than by value-based ranking.","section":"Section 'CoEvo-Mem Framework', Eq. (5); Table 3"},{"comment":"The value update is self-referential in a way that is not analyzed. Equation (6) initializes a new memory's Q as the mean of the exposed memories' Q values, and Eq. (7) uses gamma times that same mean as the bootstrap target for the exposed memories. Because Q also directly enters the retrieval ranking in Eq. (5), a miscalibrated Q could silently distort rankings. No convergence or stability analysis is provided, and the paper defers theoretical guarantees to future work. Please add an analysis or an experiment that checks Q calibration or compares against a non-bootstrapped utility estimator; without this, the link between memory evolution and retrieval improvement through Q is the weakest step in the argument.","section":"Section 'Retrieval-Aligned Relational Memory Evolution', Eqs. (6)-(7)"},{"comment":"Some rows in Table 2 are imported literature values rather than matched reruns, as disclosed in B.3 ('Imported values are system-level literature references, not paired reruns'). Because the main text does not mark which rows are imported, the 'highest overall score' comparison against StructMem and other baselines may be read as a matched comparison. Add explicit markers for imported results, or restrict the headline comparison to matched runs.","section":"Appendix B.3; Table 2"}],"minor_comments":[{"comment":"The first data row of Table 3 is formatted ambiguously: 'GPQA Diamond 68.3360.00 58.33' should be separated as '68.33 60.00 58.33', and the missing delimiter makes the ablation numbers hard to read.","section":"Table 3"},{"comment":"The caption refers to purple and peach highlighting, but these colors are not visible in a monochrome rendering; please add textual markers (e.g., asterisks or bold) for trainable model-integrated methods and for CoEvo-Mem.","section":"Table 4 caption"},{"comment":"The KL divergence in Eq. (4) uses Cat(pi_t) and Cat(\\bar p^0_t), but it is not specified whether the prior used in the KL term is the smoothed prior from Eq. (3) or the raw p^0_t; please clarify the notation.","section":"Section 'Route-Specialized Query Rewriting and Retrieval', Eq. (3)-(4)"},{"comment":"The caption says 'Training success rate' while the y-axis is labeled 'Accuracy'; please make the metric name consistent, and indicate whether the plotted values are means over the five splits.","section":"Figure 3"}],"recommendation":"major_revision","confidential_remarks":"The paper's main empirical finding is likely salvageable, but the SOTA overclaim and the missing lambda_Q = 0 ablation are the key review points. If the authors can provide an ablation isolating Q and revise the abstract to match Table 4, the contribution would be much stronger."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"CoEvo-Mem is a serious engineering paper with a reproducible recipe for jointly adapting a retrieval router and a relational memory bank. The core result—that alternating co-evolution beats fixing one side—is supported by the ablations, and the appendix is unusually careful about failure semantics and hyperparameters. Worth a serious referee, but the paper oversells itself. The abstract claims SOTA across seven benchmarks; Table 4 shows the method is second to ElasticMem on both MC tasks, and the LoCoMo OpenQA comparison is not fully matched (imported baselines, category 5 removed). That needs fixing.\n\nWhat is actually new: the combination of route-specialized rewriting, a residual REINFORCE router that corrects an LLM prior, a Q-weighted hybrid retriever, and relational TD credit propagation with alternating phases. Each ingredient has antecedents, but the closed-loop integration is original and clearly described. The experiments span seven task families, which is a lot of work, and the ablations show both main components matter: removing SR-QR or memory evolution drops macro-average by 4-5 points, and alternating beats simultaneous and two-stage.\n\nWhere the evidence is soft: the learned memory utility Q is never isolated. The 'w/o Memory Evolution' ablation removes typed relations, temporal expansion, and graph propagation together, so the 4.80-point drop cannot be attributed to Q. No experiment sets lambda_Q=0 while keeping the graph. That matters because the Q update is self-referential: Equation (6) initializes a new memory's value as the mean of the exposed set's Q, and Equation (7) uses gamma times that same mean as the bootstrap target for the exposed memories. Without a stability analysis, a miscalibrated Q could rank irrelevant memories above relevant ones; the system might still work because lambda_Q is small (0.15) and dense/sparse RRF dominates. The claim that memory evolution improves ranking is therefore untested as stated.\n\nMinor but real: a dozen hyperparameters (lambda_Q, gamma, alpha_Q, rho's, etc.) are set without sensitivity analysis. That is acceptable in a first system paper but should be acknowledged.\n\nWho this is for: anyone working on long-term memory for LLM agents. The paper gives a concrete, high-performing recipe with enough detail to reimplement. I would cite it as a system paper, not as evidence for a specific utility-learning mechanism.\n\nRecommendation: engage with it—send it to peer review. But the authors need to either add an ablation where the Q-term is neutralized (lambda_Q=0 or all values equal) or soften the memory-evolution claim. Expect a conditional accept after revision.","headline":"Careful, reproducible system paper; the co-evolution result holds up but the SOTA claim and the un-isolated Q component need attention.","tokens_in":22512,"tokens_out":4727,"would_cite":true,"duration_ms":38608,"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":"The paper claims that long-term LLM agents should co-evolve the retrieval policy and the memory bank, because retrieval determines which memories receive credit while memory updates determine what later retrieval can find.","keywords":["long-term memory","LLM agents","retrieval-augmented generation","memory utility learning","relational memory graph","query routing","co-evolving retrieval and memory","alternating optimization"],"falsifier":"Freeze the router, graph structure, and query rewrites from a trained run, replace every learned utility $Q(m)$ with a random permutation of the same values, and re-run the held-out evaluation. If performance does not drop by roughly the margin the paper attributes to memory evolution, the learned utility signal is not carrying the claimed load.","tokens_in":21487,"feed_emoji":"🧠","tokens_out":11187,"duration_ms":91207,"temperature":0.7,"pith_summary":"Long-term LLM agents need both to find the right past experiences and to keep those experiences useful. The paper's central claim is that these are not two separate problems: retrieval decides which memories receive feedback, and updates to memory values and relations change what later retrieval finds. It proposes CoEvo-Mem, which closes that loop by coupling a lightweight residual router to a typed relational memory graph, and alternates router updates with memory evolution to keep the pair stable. The paper reports the highest mean scores in all six Table 1 settings, gains up to 7.50 points over the strongest baseline, and first place on LoCoMo free-form QA, supporting the claim that retrieval and memory should co-evolve.","feed_headline":"Co-evolving retrieval and memory lifts scores by 7.5 points","feed_subtitle":"A closed feedback loop between retrieval and memory evolution beats separate tuning in long-term agent benchmarks.","key_machinery":"The carrying object is a typed relational memory graph $G_t = (V_t,E_t)$ whose nodes store experience and a learned utility estimate $Q_t(m)$, coupled to a Q-weighted hybrid retriever. Retrieval fuses dense embedding similarity, BM25 sparse lexical matching, and the utility rank through reciprocal rank fusion, a standard rank-combination formula, with route weights produced by a residual router that corrects a frozen LLM's routing prior. On a memory-phase interaction, an exposed memory receives a temporal-difference-style residual $\\delta_t(m) = R_t u_t(m) + \\gamma \\hat{Q}_t(m^{\\mathrm{new}}_t) - Q_t(m)$, where $u_t(m)$ is the attributed contribution score and $\\hat{Q}_t(m^{\\mathrm{new}}_t)$ inherits the mean utility of the exposed set; credit then propagates across paths of typed edges with relation-specific attenuation, up to depth $D$. The retrieved exposure set $A_t$ is the only meeting point of the two learning processes, and alternating router passes with memory passes is the stabilizer that makes the coupled loop trainable.","core_discovery":"The central claim is that memory access and memory evolution reinforce each other and should be optimized jointly. In CoEvo-Mem, the retrieved set is the coupling interface: a residual router, initialized to reproduce a frozen LLM's dense/sparse routing prior, is corrected online by task reward, and the same retrieved memories receive a temporal-difference-style utility update whose bootstrap value comes from the newly created memory. Typed dense, sparse, and temporal relations propagate the resulting outcome credit to neighboring memories, which changes the Q-valued ranking used by the hybrid retriever for later queries. Phase-wise alternating passes freeze one component while the other learns, limiting the non-stationarity that coupled updates create. Across seven benchmarks the paper reports top results, with improvements up to 7.50 points over the strongest baseline.","pith_inferences":["Because the router is initialized to copy the frozen LLM's prior, the marginal value of the learned router could be tested directly by comparing the full system against the same system with the router ablated while keeping the graph and query rewrites intact; the paper's SR-QR ablation approximates this but also removes the route rewrites.","The bootstrap target is self-referential, so a direct stress test of whether memory evolution comes from the bootstrap is to run with $\\gamma = 0$ and a constant inherited utility; the paper does not report this variant.","The same retrieved-set coupling should extend to multi-agent shared memory, where private routers read from and write to one graph and cross-agent credit is assigned through exposure; the paper lists this as future work."],"forward_implications":["Separately optimizing retrieval and memory leaves a measurable gap; long-term agent memory systems should treat the two as one coupled learning problem.","Only a small residual router is learned; the answering LLM, the query rewriter, and the memory models stay frozen, so the approach does not require retraining a large language model.","Task rewards can be converted into per-memory utility updates by temporal-difference-style bootstrapping, with newly distilled memories inheriting value from the context that produced them.","Alternating one full pass per component is a more effective coordination schedule than updating both components simultaneously or optimizing them in two separate stages, based on the paper's ablations."],"supporting_citations":[{"why":"It supplies MemQ, the memory-utility learning baseline that CoEvo-Mem must outperform to show co-evolution beats value-only memory evolution.","marker":"Liao et al. 2026"},{"why":"It supplies MemRL, the runtime-reinforcement-learning memory baseline compared in Table 1.","marker":"Zhang et al. 2026"},{"why":"It supplies ElasticMem and the MC-10 protocol that turns LoCoMo and LongMemEval into fixed-memory retrieval benchmarks.","marker":"Feng et al. 2026"},{"why":"It supplies StructMem, the strongest structured-memory baseline on LoCoMo free-form QA that CoEvo-Mem exceeds.","marker":"Xu et al. 2026a"},{"why":"It supplies reciprocal rank fusion, the rank-combination formula at the core of the Q-weighted hybrid retriever.","marker":"Cormack, Clarke, and Buettcher 2009"},{"why":"It supplies the score-function estimator that pushes terminal task reward through the non-differentiable retrieval step to the router.","marker":"Williams 1992"},{"why":"It supplies the REINFORCE-style optimization recipe used in the router loss.","marker":"Ahmadian et al. 2024"},{"why":"It supplies HyDE, the query-rewriting baseline that motivates the route-specific dense rewrites used by SR-QR.","marker":"Gao et al. 2023"}],"fun_headline_variants":["Closed-loop co-evolution of retrieval and memory","Retrieval and memory co-evolve for better agent recall","CoEvo-Mem: joint evolution of retrieval and memory","Co-evolving retrieval policy and memory bank for agents","Closed loop: retrieval and memory shape each other"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The system's learned memory scores are reliable enough to rank retrieval even though each new score starts as the average of the scores of the memories that produced it and that same average then feeds the target those scores learn toward; the paper provides no convergence or stability analysis for this self-referential update.","fun_headline_variants_meta":{"raw":{"variants":["Closed-loop co-evolution of retrieval and memory","Retrieval and memory co-evolve for better agent recall","CoEvo-Mem: joint evolution of retrieval and memory","Co-evolving retrieval policy and memory bank for agents","Closed loop: retrieval and memory shape each other"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001761,"raw_usage":{"total_tokens":6941,"prompt_tokens":929,"completion_tokens":6012,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":545,"completion_tokens_details":{"reasoning_tokens":5934}},"tokens_in":545,"tokens_out":6012,"duration_ms":31877,"temperature":1.0,"reasoning_tokens":5934,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T15:04:37.119580+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Freeze the router, graph structure, and query rewrites from a trained run, replace every learned utility $Q(m)$ with a random permutation of the same values, and re-run the held-out evaluation. If performance does not drop by roughly the margin the paper attributes to memory evolution, the learned utility signal is not carrying the claimed load.","supporting_citations":[{"cited_title":"Backtoba- sics:RevisitingREINFORCE-styleoptimizationforlearning from human feedback in LLMs","cited_arxiv_id":null,"evidence_quote":"It supplies the REINFORCE-style optimization recipe used in the router loss."}],"review_version":2}