{"id":"70c162c7-dde7-4c19-9b70-c3eefe1de22a","arxiv_id":"2608.11701","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"low","formal_verification":"none","parameter_count":1,"one_line_summary":"A learned slot-local transform lets a frozen backbone update and retrieve a persistent memory after context reset, and using that memory to guide routing nearly doubles updated-mapping recall.","lead":"Freezing 99.96% of a neural network, a 12.35K-parameter Consolidator module converts short-term phase memory into persistent long-term memory, and feeding that long-term memory back into routing raises updated-mapping recall from 44% to 87% on a two-segment arithmetic task. The study shows that retained memory can do more than store values: it can steer which memory slots later inputs use.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The routing ablation (Table 4) is not a same-checkpoint comparison: routing-on and routing-off each train their own Consolidator, so the 42.64 pp gain conflates the direct LTM routing term with Consolidator co-adaptation.","rationale":"The reader's weakest assumption concerns whether a shared slot-local transform can extract an update from the frozen backbone's STM. That is a representational premise, and the paper provides some supporting evidence (immediate STM recall of 89.90% in the Consolidator-only setting, plus mismatched and fresh-LTM controls). The more load-bearing issue I see is causal identification of the access-state claim. The routing-on/routing-off comparison trains separate Consolidators, so the 42.64 pp difference is not a same-checkpoint estimate of the direct LTM routing term. The paper uses a same-checkpoint method for the identity control but omits it for the routing ablation, which is the only evidence for the 'access state' half of the central claim. This is a concrete, fixable gap: a same-checkpoint toggle of Equation (7) would settle whether the direct LTM input to the router at inference is what drives the gain. Because the paper is otherwise carefully controlled and the identity-vs-learned result is robust, this concern does not overturn the verdict; it reinforces the need for the CONDITIONAL accept already given, with the added condition of reporting a same-checkpoint routing ablation. I therefore leave the verdict unchanged while flagging a different weakest point than the reader did.","tokens_in":12551,"tokens_out":8620,"duration_ms":96735,"concrete_test":"Take the trained routing-on Consolidator-only checkpoint and evaluate it on the fixed test stream with the direct LTM term removed from Equation (7), keeping all other forward paths identical. Independently, take the trained routing-off checkpoint and evaluate it with the direct LTM term added. If the routing-on checkpoint falls from 87% to roughly 44% when routing is disabled, or the routing-off checkpoint remains near 44% when routing is enabled, then the current paired ablation does not establish the access-state mechanism; if recall stays near 87% without routing, the gain is driven by Consolidator representation, not by the routing connection.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that consolidated LTM functions as an access state rests on the paired routing ablation in Section 7.2 and Table 4: direct LTM-conditioned routing raises updated-mapping recall from 44.38% to 87.02%. However, the routing-on and routing-off conditions train separate Consolidators from the same Phase-1 checkpoint. The 42.64 pp difference therefore conflates two effects: (i) the direct LTM term in Equation (7) at inference, and (ii) the fact that the routing-on Consolidator was trained with a different gradient path and may store a different LTM representation. The paper explicitly notes 'each trains its own Consolidator' and does not report a same-checkpoint toggle of the direct LTM term on a fixed Consolidator. This matters because the mechanistic statement 'retained LTM is also fed into the hierarchical router, thereby conditioning which explicit-memory slots subsequent inputs access' is about the inference-time causal role of the LTM input. The identity-vs-learned comparison in Table 3 is a clean same-checkpoint control, but the routing comparison, which is the only evidence for the access-state half of the central claim, is not. Without showing that a fixed Consolidator's recall changes when the LTM term is added or removed at evaluation, the observed gain could be entirely due to the Consolidator learning a better LTM representation under routing-on training, with the direct routing connection itself contributing little at inference. The paper's own limitation (Section 5.2) acknowledges that the task measures the combined effect on write-time and query-time routing, but that does not resolve the same-checkpoint identification problem.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces Consolidator, a 12.35K-parameter shared slot-local operator that transforms routed short-term memory (STM) into long-term memory (LTM) in a frozen Phasor Memory Network (PMNet), without replaying source tokens. After each consolidation the KV cache and STM are cleared, and the retained LTM is both readable and, in the main condition, added directly to the hierarchical router's slot-scoring state. The task is a two-segment modulo-10 mapping episode in which the second segment updates the mapping at the same address; the held-out query is answered only after both consolidations and a reset. The main same-checkpoint result is that learned consolidation gives 87.02% updated-mapping recall versus 18.32% for forced identity accumulation from the same checkpoint. The paper also reports a routing ablation in which direct LTM-conditioned routing is associated with 87.02% recall versus 44.38% without it, plus mismatched/fresh-LTM controls near chance and a dual-objective result combining pre-consolidation STM supervision with post-reset LTM recall.","tokens_in":12870,"tokens_out":3869,"duration_ms":45012,"significance":"If the central mechanistic claim is established, the paper provides a compact demonstration that persistent memory can be both retrievable content and an access state that conditions later slot selection, using only 0.041% of the model's parameters as trainable after STM pretraining. The learned-versus-identity comparison in Table 3 is a genuine same-checkpoint control, the parameter isolation is carefully described, and the limitations are explicit, including the descriptive value of n=5 statistics and the absence of long-horizon or natural-language tests. The Access-State half of the central claim, however, rests on a routing ablation that is not a same-checkpoint comparison, so the paper's main conceptual contribution is not yet fully supported.","major_comments":[{"comment":"The routing ablation that supports the claim that consolidated LTM acts as an access state is not a same-checkpoint comparison. Routing-on and routing-off each train their own Consolidator from the same Phase-1 checkpoint, so the +42.64 pp difference conflates the direct LTM term in Equation (7) at inference with differences in the trained Consolidator weights and in the LTM representation those weights produce. The text acknowledges this ('each condition trains its own Consolidator'), but the conclusion 'direct routing raises updated-mapping LTM recall' and the abstract's paired-gain wording treat the comparison as if it isolated the presence of the direct routing term. Without a same-checkpoint toggle of the LTM term in Equation (7) on a fixed Consolidator, the 42.64 pp gain could be entirely due to the routing-on Consolidator learning a better LTM representation, with the direct routing connection contributing little at inference.","section":"Section 7.2 and Table 4"},{"comment":"The claim that 'the routing ablation removes only the direct Lb,g,j term in Equation (7)' is incomplete as a statement about the intervention. In the reported experiments, the removal also changes the training gradient path and therefore the learned Consolidator itself. A reader cannot tell from the current data whether the recall difference reflects the inference-time role of the LTM input to the router or merely a different learned consolidation function. The paper should either provide the same-checkpoint evaluation (e.g., take a trained routing-on checkpoint and evaluate with and without the LTM term, or take the routing-off checkpoint and add the LTM term) or explicitly downgrade the access-state conclusion to an association rather than a causal role.","section":"Sections 4.3 and 6.2"}],"minor_comments":[{"comment":"The phrase 'direct LTM routing raises updated-mapping recall' should be qualified as 'in a comparison of separately trained conditions' wherever it appears, until a same-checkpoint toggle is reported.","section":"Abstract and Section 7.2"},{"comment":"The caption says the dual objective restores pre-consolidation recall 'without reducing' updated-mapping LTM recall; since the +1.88 pp change is not statistically resolved, 'without a statistically resolved change' would be more precise.","section":"Figure 4 caption"},{"comment":"The mismatched-experience intervention is a good content-dependence control, but it would be clearer to state explicitly that it is evaluated on the same trained checkpoint used for the correct-experience condition.","section":"Section 5.3"}],"recommendation":"major_revision","confidential_remarks":"The paper is within scope and carefully scoped, but the access-state claim is load-bearing and currently rests on a confounded routing comparison. A same-checkpoint routing toggle is a feasible and localized addition; without it, the central abstraction-level conclusion should be softened. I would be willing to review a revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: the paper is a carefully scoped proof-of-concept that a small learned transform can convert a frozen model's short-term memory into persistent long-term memory. The same-checkpoint comparison of learned vs identity consolidation (Table 3) is clean and convincing. The weaker half is the routing ablation: routing-on and routing-off each train their own Consolidator, so the 42.64 pp gain is not a pure measure of the direct LTM term. That needs flagging.\n\nWhat's genuinely new: the slot-local gated phase transform (Eq. 5), the direct LTM-conditioned routing term (Eq. 7), and the two-segment same-address update task. The task design separates carrying state across a reset, revising a mapping at a reused address, and retrieving it after STM and KV cache are cleared. The paper is also honest about limitations—single synthetic task, n=5 descriptive stats, and code that is promised but not yet shipped.\n\nWhere it's soft: the central claim that LTM functions as an access state, not just stored content, rests on the routing ablation in Table 4. Routing-off and routing-on each train their own Consolidator from the same Phase-1 checkpoint, so the +42.64 pp difference conflates the direct LTM term with co-adaptation of the Consolidator. The paper notes this in Section 6.2 (\"each trains its own Consolidator\") but still writes the abstract as though the direct term is causally identified. An evaluation-time toggle—taking the routing-on checkpoint and removing the LTM term from Eq. 7 at inference—would settle it. That's a simple missing control, not a fatal flaw, but it means the \"access state\" half of the claim is over-stated as it stands.\n\nMinor: the paper says immediate STM recall remains exactly 89.90% in both routing conditions. That's surprising if routing affects writes; presumably it's measured on a trajectory without LTM routing, but it's not made explicit. Also, the learned-vs-identity gap in the routing-off condition (+21.40 pp) shows the transform helps even without direct routing, which is nice support for the content-state half.\n\nVerdict: worth referee time. The mechanism is new, the controls are mostly well designed, and the confound is easy to fix. I'd ask for the same-checkpoint routing toggle and a slightly less assertive abstract before acceptance. For a reader in memory-augmented networks or continual learning, this is a useful, if narrow, datapoint.","headline":"A clean frozen-backbone memory consolidation result with a solid identity control; the routing ablation that supports the 'access state' claim is confounded because the two routing conditions train separate Consolidators.","tokens_in":13436,"tokens_out":3446,"would_cite":false,"duration_ms":33917,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A 12,350-parameter Consolidator converts short-term memory into a long-term store that also steers later memory access.","keywords":["explicit memory","memory consolidation","short-term memory","long-term memory","hierarchical routing","recurrent state","phasor memory","parameter-efficient adaptation"],"falsifier":"Measure the distribution of slots selected by the frozen router on the final held-out query across the 1K test episodes, with direct LTM routing on and off; if the two distributions are identical while updated-mapping recall differs by 42.64 points, then the LTM term is not acting through slot selection and the access-state reading of the result would be unsupported.","tokens_in":12370,"feed_emoji":"🧠","tokens_out":11913,"duration_ms":115278,"temperature":0.7,"pith_summary":"The paper argues that copying short-term memory into a long-term store is not enough; the retained state must also change how subsequent inputs select memory slots. In a Phasor Memory Network, a shared 12,350-parameter slot-local operator is trained to transform routed short-term memory into long-term memory at a boundary, without replaying source tokens. On a two-segment modulo-10 mapping task where the second segment updates the mapping at the same address, this raises post-reset recall of the updated mapping to 87.02%, versus 18.32% for raw accumulation and 44.38% when the same long-term memory is available only through the read path. The paper's claim is that, on this task, consolidated long-term memory acts as both retrievable content and an access state that guides later slot selection.","feed_headline":"A 12,350-parameter add-on raises frozen-model recall to 87%","feed_subtitle":"Training only a 12,350-parameter transform lifts updated-mapping recall from 18.3% to 87.0% after a reset.","key_machinery":"The central object is Consolidator, a shared slot-local gated phase transform $C_\\psi(S) = \\operatorname{atan2}(s_\\psi, c_\\psi)$ applied to occupied short-term memory slots before their values are accumulated into long-term memory by phase wrapping. Its parameters are initialized so the transform starts as identity. The second load-bearing piece is direct LTM-conditioned routing: Equation (7) adds the retained long-term state $L_{b,g,j}$ to the router's candidate-slot scoring representation, so LTM participates not only in reads but in selecting which memory slots future tokens write and read. Together these close the recurrent path $S_t \\xrightarrow{C_\\psi} L_t \\to a_{t+1} \\to S_{t+1}$ with only non-parametric memory changing after training.","core_discovery":"The paper's central claim is that a separately trained consolidation operator can make a frozen memory network persist and revise information across a reset, and that the retained long-term memory can then feed the router directly to condition which memory slots later inputs access. With the 29.95M-parameter backbone and memory interface frozen, training only the Consolidator's 12.35K parameters produces 87.02% updated-mapping LTM recall; forcing the same checkpoint to accumulate raw short-term memory instead yields 18.32%, and removing only the direct LTM term from routing drops recall to 44.38% while immediate STM recall stays at 89.90%. Mismatched experience collapses recall to 9.30%, showing the consolidated state stores episode-specific function parameters. The paper presents this as a mechanism-level proof of concept, not a general long-term memory system.","pith_inferences":["Editorial inference: because the Consolidator is shared across every address, block, and slot, a fixed trained Consolidator should in principle consolidate unseen same-family mappings without retraining, provided the Phase-1 STM representation remains intact; this is directly testable on held-out function parameters.","Editorial inference: the 42.64-point routing gain could be studied as a content/access separation: training a variant that feeds LTM to the router but disables the LTM read path would tell whether slot selection alone can carry the episode information, or whether retrievable content is also required.","Editorial inference: the dual-objective result suggests the low immediate-STM recall in the main condition was a training-signal artifact rather than an architectural ceiling; if so, scaling to longer horizons may be a scheduling problem between current-state and consolidated-state objectives, not a capacity problem."],"forward_implications":["Persistence alone is not adaptation: raw identity accumulation carries the first mapping across a reset but fails to revise it at a reused address, while the learned transform reaches 87.02%.","Retained LTM can steer slot selection in a frozen router: removing the direct LTM routing term costs 42.64 percentage points even though content and read paths are unchanged.","Learned consolidation is content-sensitive: replacing the episode's experience with a same-address, same-family donor drops recall to 9.30%, near the 10% chance level.","Immediate STM recall and post-reset LTM recall can coexist under one parameter set when the training objective supervises both, reaching 95.76% and 95.58% respectively in the dual-objective condition."],"supporting_citations":[{"why":"Supplies the Phasor Memory Network's hierarchical phase-routing architecture and memory read/write formalism that the paper freezes and extends with an STM-LTM boundary and direct LTM routing.","marker":"[8]"},{"why":"Establishes end-to-end learned addressing over external memory, the paradigm this paper positions as an alternative to append-only transformer context.","marker":"[2]"},{"why":"Extends differentiable external memory with content-based addressing and dynamic slots, giving the memory interface family the Consolidator operator is designed for.","marker":"[3]"},{"why":"Presents test-time memorization via a deep memory module; the paper contrasts its own forward-only consolidation with this adaptive-state family.","marker":"[7]"},{"why":"Motivates separate fast and slow learning stores, the computational analogy underlying the STM/LTM split and the consolidation boundary.","marker":"[17]"}],"fun_headline_variants":["Consolidator: 12K params lift frozen net to 87% recall","Tiny add-on revives memory: 87% recall after reset","Routed consolidation beats identity by 68 points (87% vs 18%)","12K trainable params: frozen model learns to remember"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The result depends on Phase-1 pretraining forming short-term state that encodes the current mapping in a way a single shared slot-local transform can read out, and the 87% recall is post-hoc evidence for that premise rather than an independent guarantee.","fun_headline_variants_meta":{"raw":{"variants":["Consolidator: 12K params lift frozen net to 87% recall","Tiny add-on revives memory: 87% recall after reset","Routed consolidation beats identity by 68 points (87% vs 18%)","12K trainable params: frozen model learns to remember"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000507,"raw_usage":{"total_tokens":2526,"prompt_tokens":1051,"completion_tokens":1475,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":667,"completion_tokens_details":{"reasoning_tokens":1395}},"tokens_in":667,"tokens_out":1475,"duration_ms":13141,"temperature":1.0,"reasoning_tokens":1395,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T00:30:40.034135+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Measure the distribution of slots selected by the frozen router on the final held-out query across the 1K test episodes, with direct LTM routing on and off; if the two distributions are identical while updated-mapping recall differs by 42.64 points, then the LTM term is not acting through slot selection and the access-state reading of the result would be unsupported.","supporting_citations":[{"cited_title":"Phasor Memory Networks: Stable Backpropagation Through Time for Scalable Explicit Memory","cited_arxiv_id":"2605.13370","evidence_quote":"Supplies the Phasor Memory Network's hierarchical phase-routing architecture and memory read/write formalism that the paper freezes and extends with an STM-LTM boundary and direct LTM routing."},{"cited_title":"Hybrid computing using a neural network with dynamic external memory.Nature, 538:471–476,","cited_arxiv_id":null,"evidence_quote":"Extends differentiable external memory with content-based addressing and dynamic slots, giving the memory interface family the Consolidator operator is designed for."}],"review_version":1}