{"id":"38cdc824-bc3b-49a2-9e58-75cf84946b87","arxiv_id":"2608.11654","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Memory is formalized as an event basis, knowledge as its span, and optimal memory as a capacity-constrained coverage maximizer tracing a utility-capacity frontier.","lead":"This paper proposes a formal definition of agent memory as a selected set of events whose generated knowledge answers queries, with optimal memory defined by expected query coverage. It also frames memory writing as a sequential decision problem and demonstrates the definitions on a small hand-built example from Homer's Odyssey.","discovery_kind":"unification","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Single-item support (Assumption 2.1) is the load-bearing condition: if agents answer real queries by combining multiple knowledge items, coverage utility measures the wrong objective and the optimal-memory/frontier theory does not apply.","rationale":"The reader's weakest assumption identifies exactly the condition on which the paper's central reduction depends, and I agree with it, sharpening only the mechanism: because all composition is pushed into the abstract operator Phi, the validity of single-item support is not a property of queries alone but also of the particular generation operator used. The paper is unusually candid, flagging the same limitation in Section 8.2 and noting in Remark B.3 that real LLMs are at best locally monotone, so the critique is not that the proofs are wrong or that the authors hide a flaw; the formal results are internally sound and the proofs in the appendices are consistent. The issue is that the practical claim, that memory quality can be measured and optimized via expected coverage, is conditional on an assumption that is likely false for real multi-hop agents and that has not been tested anywhere in the paper. The one numerical illustration is hand-constructed, so it cannot demonstrate that the condition holds for real systems; it only shows the machinery runs when all objects are chosen by the authors. Because the reader's verdict is already CONDITIONAL and my concern is the same one, the verdict should remain unchanged: the paper should be accepted only with empirical validation of single-item support on a real system, or with a clear relaxation of the assumption, for example to a joint-answer model. The proposed check would settle the condition directly by measuring whether coverage utility and end-to-end answerability diverge on composite queries under a concrete Phi.","tokens_in":20202,"tokens_out":9978,"duration_ms":109588,"concrete_test":"On a controlled corpus with known event sets (for example, the Odyssey instance extended with multi-hop queries), instantiate Phi concretely with a fixed LLM prompt that summarizes each stored event set into knowledge items, and instantiate the reasoner as an LLM allowed to retrieve and combine items. Enumerate all memories up to capacity 3 and compute, for each query, the coverage utility u(q,M) using the LLM-generated Phi spans, and the end-to-end success of the reasoner. The concern lands if a substantial fraction (say over 10%) of correctly answered queries have u=0, or if memories ranked higher by expected coverage have lower end-to-end accuracy on composite queries.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Assumption 2.1 makes answerability equivalent to coverage: u(q,M)=1 iff some single item in Phi(M) covers q. This equivalence is the hinge of the whole paper: Definition 2.5, the optimal-memory problem (Definition 3.1), the frontier (Proposition 3.2), and the maximum-coverage reduction (Proposition 3.3) all rest on it. The paper's own limitation statement (Section 8.2) concedes that the assumption 'excludes answer-time reasoning: an agent that genuinely combines items to answer a query lies outside the account.' That is not a minor edge case: modern LLM agents routinely answer multi-hop questions by composing several retrieved facts, and the paper's own Odyssey instance defines composite queries (for example, 'For how many years in total did enchantresses detain Odysseus?') that are answerable only if Phi has already materialized a composite item. Since Phi is left abstract, whether such composite items exist for a given LLM is an empirical fact, not a consequence of the framework. If Phi fails to generate the needed composite, the memory may contain all required information yet u=0; if the reasoner composes at answer time despite u=0, coverage utility understates quality. Either way, the frontier is not a valid yardstick for such agents. Monotonicity (Assumption B.2) is a second, acknowledged risk, but it mainly distorts the frontier's shape; single-item support invalidates the objective itself. This is a condition on applicability, not an internal inconsistency: the formal results follow from the stated assumptions. But the central claim's practical force, that memory quality is measurable on a common scale, has no evidence that real systems satisfy the condition.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper develops a formal framework for defining and optimizing agent memory. It models a material D by its event set ED, a generation operator Φ that maps event sets to knowledge items, and a memory MD as a subset of ED whose 'span' is Φ(MD). Under the single-item support assumption, a query is answerable exactly when some spanned knowledge item covers it, so expected query coverage becomes the objective. The optimal memory is the capacity-constrained maximizer of this coverage utility, which defines a utility–capacity frontier; under a decomposable Φ the problem reduces to weighted maximum coverage with the classical (1−1/e) greedy guarantee, while non-decomposable interactions can break submodularity. The paper then separates coverage from precision under noisy extraction, defines the water-inflation degree, organizes memorization settings into a taxonomy, and unifies them in a sequential MDP in which memory is the state, writing is the action, and query-time utility is the reward. The framework is instantiated on a hand-constructed Odyssey example and used to position existing memory systems. The appendices contain the stated proofs.","tokens_in":20522,"tokens_out":18235,"duration_ms":179749,"significance":"If the framework is accepted, it supplies a formal object for memory, a well-posed optimality criterion, and a common utility–capacity yardstick, which would be a useful conceptual contribution to the agent-memory literature. The formal propositions are simple but correctly argued, and the reduction to weighted maximum coverage in Proposition 3.3 is standard and correctly applied. The coverage-versus-precision separation and the water-inflation degree are genuinely clarifying ideas. A particular strength is that the derivation is definitional and self-contained: no empirical fitting is used, no prediction is recycled as evidence, and the paper openly states its four main limitations. The main caveat is that the applicability of the whole optimality theory hinges on single-item support, which the paper itself acknowledges to be only an idealization for real LLM agents; the framework is best read as a conditional definitional contribution rather than a demonstrated empirical account of how LLM memory works.","major_comments":[{"comment":"The example is introduced with the statement 'The extraction procedure Φ is decomposable,' yet the same section later states that 'the same events exhibit the alternative: Φ composes events that belong together' and that for the listed pairs Φ({e,e'}) strictly contains Φ({e}) ∪ Φ({e'}). Under the definition of decomposability used in Proposition 3.3(a), these two statements cannot both hold for the same operator: a decomposable Φ must satisfy Φ({e,e'}) = Φ({e}) ∪ Φ({e'}). The frontier in Table 4 and Figure 2 is computed using the decomposable operator, so the composite items introduced in this paragraph appear to belong to a different, non-decomposable variant that is never explicitly defined. Because this example is the paper's only concrete illustration and is advertised as exhibiting 'the synergy of interacting events,' the confusion is load-bearing and must be resolved before the example can support the intended reading.","section":"Section 6, 'Where events interact'"},{"comment":"Single-item support is the hinge of the entire framework: Definition 2.5, Definition 3.1, Proposition 3.2, and Proposition 3.3 all rely on the equivalence between answerability and coverage by a single knowledge item. Section 8.2 correctly concedes that 'an agent that genuinely combines items to answer a query lies outside the account.' This is a scope condition rather than an internal inconsistency, but the abstract and Section 1 present the framework as 'a unified formal account' of large-model agent memory, and Section 7 maps representative LLM systems, many of which perform multi-hop or answer-time reasoning, onto the framework without repeating this caveat. The revision should state the condition as an explicit scope boundary early in the paper, and should indicate how one would check for a given LLM whether Φ actually materializes the composite items on which the validity of the coverage objective depends.","section":"Assumption 2.1 and Section 8.2"},{"comment":"The paper optimizes coverage and adopts a 'proxy stance' that sufficient information approximately equals answerability, but the only numerical illustration is a hand-built coverage/precision computation with no reasoning policy. The limitation section says the example 'supports the stance' without establishing it; in fact the example contains no end-to-end reasoner at all, so it cannot provide even suggestive evidence for the proxy stance. Since the optimality results are explicitly about coverage rather than end-to-end correctness, the paper should consistently describe M*_D(S) as a coverage-optimal memory and should not rely on the example to justify the proxy.","section":"Section 8.2, coverage-is-not-correctness"}],"minor_comments":[{"comment":"The symbol M*_D(S) is reused in Definition 4.3 for a different objective (precision instead of coverage). Renaming the noisy version, for example M*_D,prec(S), would avoid confusion when both frontiers are discussed.","section":"Definitions 3.1 and 4.3"},{"comment":"In the Level 1a row, the 'Object of learning' column says 'optimal memory, one-shot,' but at Level 1a the query distribution is known and there is nothing to learn; 'object of computation' would be more accurate.","section":"Table 1"},{"comment":"'Homer'sOdyssey' is missing a space, and the section alternates between a decomposable Φ and a non-decomposable discussion without explicitly marking the switch; a clear statement of which operator is in force for which calculation is needed.","section":"Section 6, opening"},{"comment":"The notation E⊂N is used for the claim that every event is a degenerate knowledge item; if non-strict inclusion is intended, E⊆N would be clearer, especially since the Appendix discusses self-containment with {e}⊆Φ({e}).","section":"Assumption B.1"},{"comment":"The paper says the frontier makes 'how good is a memory' measurable, but U*_D(S) is defined relative to a specific Φ, query distribution pD, and answerability relation ans; the text should explicitly say that this is measurability relative to a formal specification, since obtaining these objects for real systems is left open.","section":"Section 3.2 and Section 7"}],"recommendation":"major_revision","confidential_remarks":"The formal core of the paper is sound and the authors are unusually explicit about limitations, which is a strength. The main risks are the internal contradiction in the only numerical example and the gap between the 'unified account' framing and the single-item support assumption that actually scopes the theory. I do not view the absence of experiments as a deficiency for this type of definitional/theoretical contribution, provided the claims are scoped accordingly. The Section 6 contradiction is the clearest blocking issue and should be fixed before acceptance."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nWhat you should know: this paper gives agent memory a clean formal skeleton—memory as a basis, knowledge as its span, optimality as capacity-constrained coverage—and it does it with honest proofs. The theorems are correct under the assumptions; the assumptions, especially single-item support, are the catch.\n\nThe genuinely new part is the mapping: a memory is a subset of events, Phi generates knowledge, and answerability is coverage by a single item. That turns optimal memory into weighted maximum coverage, gives the utility–capacity frontier, and the sequential MDP framing of writing is a natural way to think about delayed reward. The noise section, separating coverage from precision and defining 'water-inflation,' is also a nice touch. The paper cites the literature broadly and maps existing systems onto the components, which is useful for positioning.\n\nThe soft spot is exactly the one the stress-test flags. Assumption 2.1 (every query answered by a single knowledge item) is doing all the work. If an agent composes multiple items at answer time—which is common for real LLM agents—coverage utility doesn't measure what memory should optimize, and the whole optimality theory stops applying. The paper acknowledges this in Section 8.2, and it's honest, but it's not a minor edge case. It's a boundary condition on the framework's applicability. Monotonicity (B.2) is also approximate for real LLMs, though that mostly distorts the frontier's shape rather than invalidating the objective.\n\nThe other gap: the only numerical instantiation is a hand-built Odyssey example. That's fine for illustration, but the paper repeatedly claims to make memory quality 'measurable' and comparable. That claim is promissory until someone fits a real system's Phi and p_D. No code, no data, no experiments. For a framework paper that's acceptable if the authors stay modest; they mostly do, but the abstract's 'common yardstick' language oversells.\n\nWho should read it: anyone working on agent memory who wants a shared vocabulary; it's a useful conceptual anchor even if they won't adopt coverage utility as their objective. I'd send it to review—it's coherent, original, and the limitations are explicit—but I'd expect the referee to push for either empirical validation or a serious relaxation of single-item support. As is, conditional acceptance is the right verdict.","headline":"A clean, honest formalization of agent memory whose coverage-theoretic core is correct under its stated assumptions; the single-item-support assumption keeps it from being a practical yardstick yet, but it deserves serious engagement.","tokens_in":21053,"tokens_out":2493,"would_cite":true,"duration_ms":27463,"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":"This paper proposes a formal definition of agent memory: a memory is a basis of stored events, its knowledge is the span generated by a generation operator, and the optimal memory is the capacity-limited choice that maximizes expected…","keywords":["agent memory","formal definition","utility-capacity frontier","coverage utility","generation operator","sequential memorization MDP","memory precision","memory basis and span"],"falsifier":"A concrete test: find or construct a query that an agent answers correctly only by combining two separate stored items at answer time, with no single spanned knowledge item covering the query. If the end-to-end answer is correct while the coverage utility is 0, Assumption 2.1 fails for that setting and the coverage-optimality theory does not apply; likewise, observing a stored claim that lowers rather than raises end-to-end answer quality would violate the monotonicity assumption on which the frontier picture depends.","tokens_in":19988,"feed_emoji":"🧠","tokens_out":8013,"duration_ms":78983,"temperature":0.7,"pith_summary":"The paper is an attempt to give agent memory a formal definition and an optimality theory. It proposes that a memory is a basis: a subset of the events extracted from a material, and that the knowledge an agent can actually use is the span of that basis under a generation operator that turns event sets into knowledge. Under a single-item support assumption, a query is answerable exactly when one generated knowledge item covers it, so the best memory of a given size is the one that covers the most query mass. That optimality condition traces a utility–capacity frontier, giving memory systems a common yardstick and making 'how good is a memory' a measurable quantity. The same framework is then extended to noisy extraction, where coverage and precision diverge, and to sequential writing, formalized as a Markov decision process with memory as state and writing as action.","feed_headline":"Memory is a basis: best agent memory maximizes query coverage","feed_subtitle":"A utility-capacity frontier makes 'how good is a memory' measurable, separating true coverage from inflated precision.","key_machinery":"The machinery is the two-layer separation of events from knowledge together with the generation operator $\\Phi:2^E\\to 2^N$, which maps any set of events to the knowledge it entails. Memory is a basis $M_D\\subseteq E_D$; its span is $\\Phi(M_D)$; and, under Assumption 2.1's single-item support, a query is answerable exactly when it lies in the union of the answerable sets $Q(n)$ of the spanned items. This reduces memory construction to a coverage problem whose optimal value defines the utility–capacity frontier $U_D^*(S)$, the key yardstick; under decomposability it is weighted maximum coverage, with greedy attaining $1-1/e$. A second mechanism, the good span $\\Phi(M_D\\cap E_D)$, separates coverage from precision under noise, and the sequential memorization MDP with state $M_{m,t}$, write action, and query-time reward unifies writing as a learning problem.","core_discovery":"In the paper's own terms, the central claim is that agent memory is well defined as a subset $M_D\\subseteq E_D$ of a material's events; its span $R_{M_D}=\\Phi(M_D)$ is the knowledge those events generate, and the optimal memory of capacity $S$ is the constrained maximizer of expected coverage utility $u(q,M_D)=\\mathbf{1}[q\\in\\bigcup_{n\\in\\Phi(M_D)} Q(n)]$ with respect to the query distribution $p_D$. The frontier $U_D^*(S)$ attained by that maximizer is non-decreasing in $S$ and saturates at the full-context baseline, delimiting a compression zone in which memory matters. When the generation operator decomposes over events, the problem reduces to weighted maximum coverage, which is NP-hard with a greedy $1-1/e$ approximation; in general it is monotone set-function maximization where cross-event synergy defeats such guarantees. The paper further claims that under noisy extraction, coverage and precision part ways, with the water-inflation degree measuring how much apparent memory quality is bought with false claims, and that continual memorization is a sequential MDP in which delayed query-time settlement is the reward.","pith_inferences":["Beyond the paper: Because the framework's yardstick only needs a material, a query distribution, and an operator $\\Phi$, one could approximate the frontier $U_D^*(S)$ on real corpora using an LLM as $\\Phi$, producing a common ranking of memory modules without new benchmarks.","Beyond the paper: The single-item support assumption implies a measurable diagnostic: if a substantial share of correct agent responses requires composing several retrieved knowledge items at answer time, the coverage objective understates memory value and a joint-answer model is needed.","Beyond the paper: The water-inflation degree could be used as a predictor for hallucination-driven memory failure: agents whose stored claims are mostly false should show a large gap between coverage and precision, and injecting a conflicting false claim into a locally monotone $\\Phi$ should visibly reduce end-to-end correctness."],"forward_implications":["Memory quality becomes quantifiable: every write policy defines a point in the size–utility plane below the frontier, and its vertical gap to the frontier is the memory-efficiency loss.","The frontier saturates at the full-context baseline, so memory matters only in the compression zone $S<|E_D|$, where the vertical gap to that baseline is the cost of compression.","With a decomposable generation operator, optimal memory is weighted maximum coverage, so greedy selection attains the $1-1/e$ approximation; with interacting events, cross-event synergy defeats such guarantees.","Under noisy extraction, optimizing coverage can prefer false claims; precision utility and the water-inflation degree separate true quality from inflated quality, and the noisy optimal-memory problem should target precision.","Writing becomes a delayed-reward learning problem formalized as a sequential MDP, with credit assignment and trust estimation as the core difficulties of learning a write policy."],"supporting_citations":[{"why":"Provides the complementary-learning-systems view of biological memory that motivates the continual, multi-level memory formulation.","marker":"(Kumaran et al., 2016)"},{"why":"Supplies the episodic-versus-semantic memory distinction that frames memory as a structured object rather than an undefined word.","marker":"(Tulving, 1972)"},{"why":"Documents the fragmented definitions and evaluation metrics for LLM-agent memory that the framework sets out to unify.","marker":"(Du et al., 2025)"},{"why":"One of the persistent paged-store systems whose write and read components the framework places in its common mapping.","marker":"(Packer et al., 2023)"},{"why":"Reflection-based agent memory used to instantiate the write-and-read decomposition and the end-to-end layer.","marker":"(Park et al., 2023)"},{"why":"Provides the long multi-session conversation testbed cited as the closest Level 1b evaluation setting.","marker":"(Maharana et al., 2024)"},{"why":"Provides a long interactive memory benchmark for the disclosure-driven Level 1b setting.","marker":"(Wu et al., 2025)"},{"why":"Knowledge-graph memory used as the representative generation-operator choice in the system mapping.","marker":"(Jiménez Gutiérrez et al., 2024)"}],"fun_headline_variants":["Memory = basis, knowledge = span, optimality = coverage","Optimal memory: capacity-constrained maximizer of expected coverage","Noisy memory inflates coverage; precision is the casualty","Utility-capacity frontier makes memory quality measurable"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is single-item support: every query must be answerable by one generated knowledge item, with all composition pushed into the generation operator; if real queries require combining stored items at answer time, the coverage utility is not the right objective and the optimality theory collapses.","fun_headline_variants_meta":{"raw":{"variants":["Memory = basis, knowledge = span, optimality = coverage","Optimal memory: capacity-constrained maximizer of expected coverage","Noisy memory inflates coverage; precision is the casualty","Utility-capacity frontier makes memory quality measurable"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001224,"raw_usage":{"total_tokens":5099,"prompt_tokens":1079,"completion_tokens":4020,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":695,"completion_tokens_details":{"reasoning_tokens":3954}},"tokens_in":695,"tokens_out":4020,"duration_ms":31761,"temperature":1.0,"reasoning_tokens":3954,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T00:31:53.547234+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A concrete test: find or construct a query that an agent answers correctly only by combining two separate stored items at answer time, with no single spanned knowledge item covering the query. If the end-to-end answer is correct while the coverage utility is 0, Assumption 2.1 fails for that setting and the coverage-optimality theory does not apply; likewise, observing a stored claim that lowers rather than raises end-to-end answer quality would violate the monotonicity assumption on which the frontier picture depends.","supporting_citations":[{"cited_title":"McClelland","cited_arxiv_id":null,"evidence_quote":"Provides the complementary-learning-systems view of biological memory that motivates the continual, multi-level memory formulation."},{"cited_title":"Episodic and semantic memory","cited_arxiv_id":null,"evidence_quote":"Supplies the episodic-versus-semantic memory distinction that frames memory as a structured object rather than an undefined word."},{"cited_title":"O'Brien, Carrie J","cited_arxiv_id":null,"evidence_quote":"Reflection-based agent memory used to instantiate the write-and-read decomposition and the end-to-end layer."}],"review_version":1}