StateFuse: Deterministic Conflict-Preserving Memory for Multi-Agent Systems

Sergey Volkov; Yang Li; Ye Luo

REVIEW 3 major objections 6 minor 21 references

Preserving contradictions in agent memory prevents 40% false-confident actions

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · glm-5.2

2026-07-08 22:39 UTC pith:MHTTO4K6

load-bearing objection Honest, narrow contract paper with clean formal objects but thin evaluation; the one StateFuse-specific advantage rests on 13 synthetic tasks. the 3 major comments →

arxiv 2607.05844 v1 pith:MHTTO4K6 submitted 2026-07-07 cs.AI cs.CLcs.MA

StateFuse: Deterministic Conflict-Preserving Memory for Multi-Agent Systems

Sergey Volkov , Yang Li , Ye Luo This is my paper

classification cs.AI cs.CLcs.MA

keywords agent memoryconflict resolutionCRDTOpSetmulti-agent systemsconflict-preserving memorysemantic correctionabstention

verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

StateFuse defines a memory contract for multi-agent systems that keeps conflicting observations visible rather than silently overwriting them. Built on standard CRDT/OpSet merge — it introduces no new join algebra — the contract adds an agent-facing semantics layer with immutable history, explicit conflict objects, dual correction handles (an exact identifier and a semantic handle), deterministic predicate contracts, and projection-time resolution that can choose among candidates or abstain but cannot rewrite replicated state. The paper's claim is deliberately narrow: StateFuse does not improve answer accuracy over strong flat multi-value baselines. On a 282-question conflict-bearing benchmark slice, all methods tie at 97.5% accuracy (64.9% under conservative abstention). What differs is what the memory surface exposes. Conflict-preserving surfaces — including non-StateFuse baselines — surface 100% of contradictions; collapsed surfaces surface none. In a controlled agent loop with uniform verification, collapsed memory produces 40% false-confident actions and only 60% post-verification success, while every conflict-preserving surface reaches 100% success with zero false confidences. A correction-handle ablation shows that semantic handles recover corrected values in 100% of tasks where exact identifiers are unavailable, compared to 76.9% for exact IDs. The paper positions StateFuse not as an accuracy win but as a safer, more auditable public memory contract for systems where disagreement and correction are operational concerns.

Core claim

The central finding is that the safety gap between collapsed and conflict-preserving memory surfaces is large and consequential, while the accuracy gap between StateFuse and strong flat baselines is zero. Collapsed memory hides contradictions entirely and leads to 40% false-confident actions in a downstream agent loop, whereas any surface that preserves ambiguity — whether StateFuse or a flat multi-value register — reaches full post-verification success under the same verification budget. The one component where StateFuse's specific design choices matter beyond what flat baselines offer is semantic correction: the claim_ref handle, derived deterministically from claim keys and predicate-gov-

What carries the argument

StateFuse's contract layer sits on top of standard OpSet/CRDT set-union merge and adds five components: (1) immutable replicated history for evidence, claims, retractions, and decisions; (2) explicit public conflict objects emitted at projection time for functional predicates with multiple distinct active values; (3) dual correction handles — claim_id for exact local edits and claim_ref for cross-replica semantic targeting; (4) deterministic predicate contracts specifying normalization, equality, and claim-reference derivation; (5) bounded projection authority where resolvers may choose, abstain, or fail closed but cannot mutate base memory. Materialization is a pure deterministic function:它

Load-bearing premise

The evaluation assumes that a 282-question benchmark slice and a 50-task synthetic agent loop are representative of real-world agent memory conflicts. The paper acknowledges both are controlled settings. If naturally arising traces show different conflict patterns, correction frequencies, or cross-replica identifier availability, the safety gap between collapsed and conflict-preserving surfaces could shrink or vanish.

What would settle it

Show a naturally arising agent trace where collapsed memory surfaces produce false-confidence rates comparable to conflict-preserving surfaces (i.e., near 0%), or where semantic correction handles fail to recover corrected values at rates no better than exact identifiers.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

Agent memory systems that collapse conflicting observations before downstream reasoning may produce false-confident actions at rates high enough to be operationally unacceptable in safety-critical settings.
The accuracy tie between StateFuse and flat multi-value baselines suggests that the value of explicit conflict objects lies in auditability and correction expressiveness, not in better answer selection — which reframes how agent memory should be evaluated.
Semantic correction handles that remain stable across replicas even when original writer identifiers are unavailable could become a standard requirement for multi-agent memory contracts, as exact-ID-only correction fails on 23% of cross-replica correction tasks.
The matched-information evaluation protocol — giving every surface the same resolver family and verification budget — could be adopted more broadly to separate the effects of information preservation from policy advantages in agent memory benchmarks.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

If real-world agent traces exhibit conflict patterns where the latest write is not the gold answer more often than in this benchmark slice, the accuracy gap between collapsed and conflict-preserving surfaces would widen, strengthening the case for conflict preservation beyond the safety argument.
The fact that flat multi-value registers match StateFuse on all measured metrics except semantic correction raises the question of whether a simpler extension to flat registers — adding a claim_ref-like semantic handle — could capture most of StateFuse's practical benefit without the full contract machinery.
The 40% false-confidence rate for collapsed memory under uniform verification suggests that even verification budgets may be insufficient if the memory surface has already destroyed the evidence needed to detect the conflict — a structural limitation that no amount of downstream checking can recover from.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Desk Editor's Note

Honest, narrow contract paper with clean formal objects but thin evaluation; the one StateFuse-specific advantage rests on 13 synthetic tasks.

read the letter

Short version: this is a well-specified memory contract paper that knows exactly what it is and isn't claiming. The formal objects are clean, the evaluation is honest, and the central argument holds. The binding constraint is evaluation depth — the sole StateFuse-specific empirical advantage is a 13-task synthetic ablation, and the paper admits it. Deserves a serious referee who can assess whether the contract design contribution is enough on its own or whether the evaluation gap is disqualifying for the venue in question. I lean toward accepting for review because the contract layer is genuinely useful and the paper is unusually non-circular in its framing. The stress-test note is right to flag the 13-task ablation as load-bearing, but I think it slightly overstates the problem: the formal semantics in Appendix A.1 are sound and the claim_ref mechanism is well-motivated by construction, so the ablation is confirmatory rather than the only evidence. The real gap is that no natural-trace evidence exists yet, which the paper acknowledges. The reader's take is accurate and well-calibrated throughout. One area where I'd push back on the reader: the significance score of 4.0 feels slightly generous given that StateFuse ties flat multi-value baselines on every measured metric except the 13-task ablation. The contract design is a real contribution, but the empirical case for it being a *necessary* contract (rather than a reasonable one) is not yet made. The matched-information evaluation protocol is genuinely good methodology and deserves credit — keeping downstream policy budgets fixed across surfaces is the right experimental design and is not common enough in this area. The related-work positioning (Table 1) is fair and does not overclaim novelty at the CRDT layer. The proposition sketches in A.3 are straightforward but correct: convergence, deterministic materialization, retraction semantics, and projection non-interference all follow directly from the OpSet model as stated. No formal proof assistant is used, but the arguments are simple enough that this is not a concern. The GitHub repo is referenced without a commit hash, which is a minor reproducibility gap. Net: this is a solid systems design paper with a narrow but defensible claim, honest about its own limitations, with an evaluation that is adequate for the contract-design contribution but insufficient for any stronger empirical claim. The right referee is someone who can evaluate the contract as an engineering contribution and judge whether the evaluation protocol, despite being narrow, is rigorous enough to support the paper's scoped conclusions. I think it is.

Referee Report

3 major / 6 minor

Summary. This paper presents StateFuse, a conflict-aware memory contract for multi-agent systems built on standard OpSet/CRDT merge. Rather than proposing a new join algebra, StateFuse defines an agent-facing semantics layer with immutable history, explicit ConflictSet objects, dual correction handles (claim_id for exact local edits, claim_ref for semantic cross-replica correction), deterministic predicate contracts, and projection-time resolution that cannot rewrite replicated state. The evaluation compares StateFuse against flat multi-value, raw-log, provenance-style, and collapsed baselines under matched resolver and verification policies on a 282-question MemoryAgentBench conflict-bearing slice, a 50-task synthetic agent loop, and a 13-task correction-handle ablation. The central claim is deliberately narrow: StateFuse is best supported as a safer public memory contract for contradiction surfacing, abstention, and auditable correction—not as a universal accuracy gain.

Significance. The paper makes a honest and well-scoped contribution. Its strongest points are: (1) the formal objects in Appendix A.1 are clean, and the proposition sketches (convergence, deterministic materialization, retraction semantics, conflict-set soundness, projection non-interference, replayability) follow directly from the OpSet model; (2) the evaluation is notably non-circular—the paper does not claim accuracy improvements and then show them, and it explicitly reports that StateFuse ties flat multi-value baselines on all main metrics; (3) the matched-information evaluation design, where all baselines receive the same resolver family and verification budget, is a methodological improvement over typical agent-memory evaluations; (4) reproducible code is provided. The contract design—particularly the distinction between exact and semantic correction handles and the bounded projection authority—is a useful engineering contribution for the agent-memory community.

major comments (3)

Table 5 (§4.3): The sole StateFuse-specific empirical advantage rests on 13 synthetic tasks where the semantic-target subset is constructed so that exact prior identifiers are unavailable. This is precisely the condition that makes claim_ref necessary by construction. The paper acknowledges in §5.5 that 'the downstream agent loop and semantic-handle ablations are still controlled synthetic evaluations' and that no 'large naturally arising trace study' has been conducted. However, the claim of 'auditable correction' as a StateFuse-specific benefit (§5.1, contribution list item 1) is load-bearing on the practical frequency of this scenario. If real multi-agent traces typically share a common op-id namespace (making exact identifiers available across replicas), the semantic handle adds no demonstrated advantage over a flat multi-value register with exact-ID retraction. The paper should要么 (a
Table 5 (§4.3): The 3-task difference between claim_ref (13/13) and claim_id (10/13) on 13 tasks is a small sample. While the paper states runs are deterministic, a brief sensitivity analysis or at minimum an explicit acknowledgment of the sample size limitation in the discussion would strengthen the claim. The current text in §5.1 ('changes which corrections are expressible') is strong relative to the evidence base.
§4.2, Table 2 Panel B: The agent loop uses 50 synthetic tasks. The paper is transparent that this is synthetic (§4.2: 'The agent loop is still synthetic'), but the headline result that 'all conservative non-collapsing surfaces reach full post-verification success' is used to support the claim that 'preserving ambiguity enables safer abstention.' Since all non-collapsing surfaces (including non-StateFuse baselines) achieve identical results, this evidence supports the value of conflict preservation generally, not StateFuse specifically. The paper should clarify in the abstract and conclusion that the agent-loop evidence supports conflict-preserving surfaces as a class, not StateFuse's contract design in particular.

minor comments (6)

§3.2: The claim_ref derivation is described as 'a deterministic semantic handle derived from the claim key and predicate-governed value contract,' but the exact derivation function is not specified. A concrete example (e.g., hash of (namespace, subject, predicate, normalize(value))) would help readers and implementers.
Table 1: The 'Resolver Boundary' column for StateFuse says 'Immutable base / mutable projection' but the cell formatting is ambiguous—it could be read as two separate entries. Consider clarifying with a brief phrase like 'base immutable; projection mutable.'
§3.6: The 'Decision error rate' metric is defined but does not appear in any results table. Either include it in the results or note that it was not used in this evaluation.
Appendix A.3: The proposition sketches are correct but very brief. For Proposition 3 (Retraction Semantics), explicitly noting that the unseen-target no-resurrection property follows from the set-membership definition of A(O) (independent of arrival order) would make the argument self-contained.
§2.5, Table 1: The 'Det.' (Determinism) column for 'Classical KV/overwrite memory' is listed as 'No' while 'RAG and agent memory stacks' is listed as 'Limited.' This distinction could use a brief footnote explaining what 'Limited' determinism means in this context.
The abstract states 'semantic handles matter when exact prior identifiers are unavailable'—consider qualifying with 'in controlled settings' to match the paper's own scoping in §5.5.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for a careful and fair reading. The three major comments are all substantive and correct in their core observations. We agree with all three and will revise accordingly: (1) we will add an explicit discussion of the conditions under which claim_ref provides practical advantage over exact-ID retraction, including the cross-replica namespace scenario, and scope the 'auditable correction' contribution claim more carefully; (2) we will add an explicit sample-size limitation acknowledgment for the 13-task ablation and soften the language in §5.1; (3) we will clarify in the abstract and conclusion that the agent-loop evidence supports conflict-preserving surfaces as a class, not StateFuse's contract design specifically. The referee's recommendation of minor revision is appropriate.

read point-by-point responses

Referee: Table 5 (§4.3): The sole StateFuse-specific empirical advantage rests on 13 synthetic tasks where the semantic-target subset is constructed so that exact prior identifiers are unavailable. This is precisely the condition that makes claim_ref necessary by construction. The claim of 'auditable correction' as a StateFuse-specific benefit (§5.1, contribution list item 1) is load-bearing on the practical frequency of this scenario. If real multi-agent traces typically share a common op-id namespace, the semantic handle adds no demonstrated advantage over a flat multi-value register with exact-ID retraction.

Authors: The referee is correct on both points. First, the semantic-target tasks are constructed so that exact identifiers are unavailable, which is indeed the condition that makes claim_ref necessary by construction. We should have been more explicit about this structural fact and about what it implies for the generality of the claim. Second, the practical value of claim_ref does depend on how often cross-replica correction scenarios arise in real traces where the correcting replica lacks the original writer's opaque op-id. We do not currently have evidence on this frequency, and the paper should say so directly rather than letting the contribution-list language ('auditable correction') imply broader support than the evidence provides. We will revise as follows: (a) add a paragraph in §5.1 explicitly stating that the claim_ref advantage is demonstrated only under the condition where exact identifiers are unavailable across replicas, and that the practical frequency of this scenario in naturally arising traces is an open question; (b) reframe contribution item 1 to distinguish between the contract-level design (StateFuse exposes both handle types as part of its public contract) and the empirical evidence (the ablation shows claim_ref changes expressibility only under the identifier-unavailable condition); (c) add to §5.5 an explicit note that if real multi-agent deployments typically share a common op-id namespace, the marginal value of claim_ref over exact-ID retraction would be reduced. We agree this is a genuine limitation of the current evidence and should be stated as such. revision: yes
Referee: Table 5 (§4.3): The 3-task difference between claim_ref (13/13) and claim_id (10/13) on 13 tasks is a small sample. While the paper states runs are deterministic, a brief sensitivity analysis or at minimum an explicit acknowledgment of the sample size limitation in the discussion would strengthen the claim. The current text in §5.1 ('changes which corrections are expressible') is strong relative to the evidence base.

Authors: The referee is correct that 13 tasks is a small sample and that the language in §5.1 is stronger than the evidence base warrants. Although the runs are deterministic (so the 3-task gap is not a sampling artifact), the small absolute number of tasks means the result could look different with a different or larger task set. We will make two changes: (a) add an explicit sample-size limitation statement in the §4.3 discussion and in §5.5, noting that the 13-task ablation is small and that the 3-task gap should be interpreted as a proof-of-concept demonstration that claim_ref expands the set of expressible corrections, not as a reliable effect-size estimate; (b) soften the language in §5.1 from 'changes which corrections are expressible' to something like 'can expand the set of expressible corrections in the identifier-unavailable condition, though the current evidence base is limited to a small synthetic ablation.' We agree that the current phrasing overstates what 13 tasks can support. revision: yes
Referee: §4.2, Table 2 Panel B: The agent loop uses 50 synthetic tasks. Since all non-collapsing surfaces (including non-StateFuse baselines) achieve identical results, this evidence supports the value of conflict preservation generally, not StateFuse specifically. The paper should clarify in the abstract and conclusion that the agent-loop evidence supports conflict-preserving surfaces as a class, not StateFuse's contract design in particular.

Authors: The referee is correct. The agent-loop result shows that all conservative non-collapsing surfaces reach the same safe operating point, which supports the value of conflict preservation as a design property but does not distinguish StateFuse's contract from a flat multi-value register with conservative abstention. The paper already acknowledges this in §4.2 ('It does not distinguish StateFuse from strong flat conservative baselines on this controlled task family'), but the abstract and conclusion do not carry this qualification forward. We will revise the abstract to clarify that the agent-loop evidence supports conflict-preserving surfaces as a class, and we will add a corresponding clarification in the conclusion. Specifically, the abstract sentence 'In a controlled agent loop with uniform verification, preserving ambiguity enables safer abstention and correction than early collapse' will be revised to make clear that this result holds for all non-collapsing surfaces tested, not for StateFuse specifically. The conclusion will include a parallel statement. revision: yes

Circularity Check

0 steps flagged

No significant circularity; the paper is self-contained and honest about what its evidence shows.

full rationale

The paper's derivation chain is largely self-contained and does not exhibit the circular patterns this analysis targets. (1) The central empirical claim—that StateFuse ties flat multi-value baselines on accuracy but differs on conflict surfacing—is evaluated against an external benchmark (MemoryAgentBench) with matched baselines, and the paper explicitly states it does not claim an accuracy advantage. (2) The correction-handle ablation (Table 5) tests whether claim_ref recovers corrections that claim_id cannot. While the ablation is constructed so that exact identifiers are unavailable on semantic-target tasks, the paper frames this as a test of expressibility ('the semantic handle changes what corrections are expressible rather than merely renaming an exact retraction'), not as an empirical prediction forced by a fit. The formal semantics (Appendix A.1, Proposition 3) independently establish that retraction by ref(c) deactivates any claim with that ref regardless of arrival order; the ablation confirms the implementation matches the specification. This is verification, not circularity. (3) The CRDT/OpSet foundations are cited to standard external work (Kleppmann, Shapiro, Almeida et al.) and the paper explicitly disclaims a new join algebra. (4) The one self-citation is to AgentGit [8] (Li and Luo are co-authors), but it is described as 'complementary' and is not load-bearing for any central claim. The paper's honesty about its limitations (§5.5: synthetic evaluations, single benchmark slice, no natural-trace study) further reduces circularity risk. The only mild concern is that 'conflict-preserving surfaces expose contradictions while collapsed surfaces do not' is close to definitional, but the paper does not present this as a deep empirical finding—it frames it as the design choice being evaluated for downstream consequences. Score 1 reflects the minor non-load-bearing self-citation.

Axiom & Free-Parameter Ledger

0 free parameters · 5 axioms · 3 invented entities

The axiom ledger is lean. No free parameters are fitted to data. The axioms are primarily domain assumptions about benchmark representativeness and determinism of predicate functions. The invented entities (claim_ref, ConflictSet, predicate registry) are the paper's core contributions; two of three have independent empirical evidence from ablations. The predicate registry lacks evidence beyond the evaluated predicates.

axioms (5)

standard math CRDT/OpSet set-union merge provides deterministic convergence under benign replicas
§2.1, §3.3: merge(O1, O2) = O1 ∪ O2. Standard result from Shapiro et al. 2011 and Kleppmann et al. 2018.
domain assumption Predicate normalization and equality functions are deterministic and replica-invariant
§3.2: 'The admissibility rule is simple: these functions must be deterministic and replica-invariant.' This is a design constraint, not proven. The contract checker validates on sample inputs but cannot guarantee stability on all inputs.
domain assumption The 282-question MemoryAgentBench conflict-bearing slice is representative of real agent memory conflicts
§4.2, §5.5: The main empirical evidence rests on this slice. §5.5 acknowledges 'The main external evidence is one official benchmark slice rather than a broad benchmark portfolio.'
domain assumption The synthetic 50-task agent loop generalizes to real agent deployments
§4.2: 'The agent loop is still synthetic.' §5.5: 'The downstream agent loop and semantic-handle ablations are still controlled synthetic evaluations.'
domain assumption LLM-backed resolvers produce valid structured JSON for replayability when responses are fixed
§A.3, Proposition 6 sketch: 'For LLM-backed resolvers, replayability holds only when the resolver response itself is fixed and validated.' This is assumed, not guaranteed.

invented entities (3)

claim_ref (semantic claim handle) independent evidence
purpose: Stable cross-replica correction handle derived from claim key and predicate-governed value contract, usable when exact claim_id is unavailable
Table 5 ablation shows claim_ref recovers corrected values in 100% of semantic-target tasks where claim_id achieves 0%. This is a falsifiable prediction tested empirically.
ConflictSet (explicit conflict object) independent evidence
purpose: Public first-class memory object representing multiple distinct active values for a functional key
Conflict recall metric (§3.6) measures whether conflicts are surfaced. Table 2 shows 100% recall for StateFuse vs. 0% for collapsed surfaces. The conflict object is observable in the projection output.
Predicate contract registry no independent evidence
purpose: Deterministic specification of normalization, equality, and claim_ref derivation per predicate
The registry is a design artifact. The contract checker (§3.2) validates on sample inputs but there is no independent evidence that the registry approach is stable across all possible predicates or that it generalizes beyond the evaluated set.

pith-pipeline@v1.1.0-glm · 14704 in / 3493 out tokens · 306823 ms · 2026-07-08T22:39:47.240574+00:00 · methodology

0 comments

read the original abstract

Agent systems accumulate conflicting observations across branches, retries, and replicas, yet many practical memory layers still collapse disagreement behind overwrite rules that are difficult to inspect or correct. We present StateFuse, a conflict-aware replicated memory contract built on standard OpSet/CRDT merge. StateFuse does not introduce a new join algebra; it defines an agent-facing semantics layer with immutable history, explicit conflict objects, exact and semantic correction handles (claim_id / claim_ref), deterministic predicate contracts, and projection-time resolution that cannot rewrite replicated state. We evaluate StateFuse against flat multi-value, raw-log, provenance-style, and collapsed baselines under matched resolver and verification policies. On a 282-question official conflict-bearing MemoryAgentBench slice, the compared methods tie on answer accuracy, but conflict-preserving surfaces keep contradictions visible while collapsed surfaces do not. In a controlled agent loop with uniform verification, preserving ambiguity enables safer abstention and correction than early collapse. A correction-handle ablation further shows that semantic handles matter when exact prior identifiers are unavailable. The resulting claim is narrow: StateFuse is best supported as a safer public memory contract for contradiction surfacing, abstention, and auditable correction, not as a universal accuracy gain.

Figures

Figures reproduced from arXiv: 2607.05844 by Sergey Volkov, Yang Li, Ye Luo.

**Figure 2.** Figure 2: StateFuse architecture: immutable merge substrate, explicit conflict materialization, and projection [PITH_FULL_IMAGE:figures/full_fig_p008_2.png] view at source ↗

**Figure 3.** Figure 3: Branch-diverge-merge sequence with conflict-preserving materialization and projection-time [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗

discussion (0)

Reference graph

Works this paper leans on

21 extracted references · 21 canonical work pages · 10 internal anchors

[1]

Delta State Replicated Data Types

Paulo S´ergio Almeida, Ali Shoker, and Carlos Baquero. Delta state replicated data types.Journal of Parallel and Distributed Computing, 111:162–173, 2018. doi: 10.1016/j.jpdc.2017.08.003. URL https://arxiv.org/abs/1603.01529

work page internal anchor Pith review Pith/arXiv arXiv doi:10.1016/j.jpdc.2017.08.003 2018
[2]

Evaluating Memory in LLM Agents via Incremental Multi-Turn Interactions

Yuanzhe Hu, Yu Wang, and Julian McAuley. Evaluating memory in LLM agents via incremental multi-turn interactions.arXiv preprint arXiv:2507.05257, 2025. URL https://arxiv.org/abs/ 2507.05257. Introduces MemoryAgentBench

work page internal anchor Pith review Pith/arXiv arXiv 2025
[3]

AMC: Towards trustworthy and explorable CRDT applications with the Automerge model checker

Andrew Jeffery and Richard Mortier. AMC: Towards trustworthy and explorable CRDT applications with the Automerge model checker. InWorkshop on Principles and Practice of Consistency for Distributed Data (PaPoC). ACM, 2023. doi: 10.1145/3578358.3591326. URL https://doi.org/ 10.1145/3578358.3591326

work page doi:10.1145/3578358.3591326 2023
[4]

Beresford

Martin Kleppmann and Alastair R. Beresford. A conflict-free replicated JSON datatype.IEEE Transac- tions on Parallel and Distributed Systems, 28(10):2733–2746, 2017. doi: 10.1109/TPDS.2017.2697382. URLhttps://doi.org/10.1109/TPDS.2017.2697382

work page doi:10.1109/tpds.2017.2697382 2017
[6]

URLhttps://arxiv.org/abs/1805.04263

work page internal anchor Pith review Pith/arXiv arXiv
[7]

Local-first software: You own your data, in spite of the cloud

Martin Kleppmann, Adam Wiggins, Peter van Hardenberg, and Mark McGranaghan. Local-first software: You own your data, in spite of the cloud. InProceedings of the 2019 ACM SIGPLAN International Symposium on New Ideas, New Paradigms, and Reflections on Programming and Software, pages 154–178, 2019. doi: 10.1145/3359591.3359737. URL https://martin.kleppmann.c...

work page doi:10.1145/3359591.3359737 2019
[8]

Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks

Patrick Lewis, Ethan Perez, Aleksandra Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich K¨uttler, Mike Lewis, Wen-tau Yih, Tim Rockt¨aschel, Sebastian Riedel, and Douwe Kiela. Retrieval-augmented generation for knowledge-intensive NLP tasks.Advances in Neural Information Processing Systems, 33:9459–9474, 2020. URLhttps://arxiv.org/abs/2005.11401

work page internal anchor Pith review Pith/arXiv arXiv 2020
[9]

AgentGit: A version control framework for reliable and scalable LLM-powered multi-agent systems.arXiv preprint arXiv:2511.00628, 2025

Yang Li, Siqi Ping, Xiyu Chen, Xiaojian Qi, Zigan Wang, Ye Luo, and Xiaowei Zhang. AgentGit: A version control framework for reliable and scalable LLM-powered multi-agent systems.arXiv preprint arXiv:2511.00628, 2025. URLhttps://arxiv.org/abs/2511.00628

work page arXiv 2025
[10]

GAIA: a benchmark for General AI Assistants

Gr´egoire Mialon, Cl´ementine Fourrier, Craig Swift, Thomas Wolf, Yann LeCun, and Thomas Scialom. GAIA: A benchmark for general AI assistants. InICLR 2024, 2024. URL https://arxiv.org/ abs/2311.12983

work page internal anchor Pith review Pith/arXiv arXiv 2024
[11]

PROV-DM: The PROV data model

Luc Moreau and Paolo Missier. PROV-DM: The PROV data model. W3C Recommendation, 2013. URLhttps://www.w3.org/TR/prov-dm/

work page 2013
[12]

MemGPT: Towards LLMs as Operating Systems

Charles Packer, Vivian Fang, Shishir Patil, Kevin Lin, Joseph E. Gonzalez, and Christopher R ´e. MemGPT: Towards LLMs as operating systems.arXiv preprint arXiv:2310.08560, 2023. URL https://arxiv.org/abs/2310.08560

work page internal anchor Pith review Pith/arXiv arXiv 2023
[13]

Bernstein

Joon Sung Park, Joseph O’Brien, Carrie J. Cai, Meredith Ringel Morris, Percy Liang, and Michael S. Bernstein. Generative agents: Interactive simulacra of human behavior.Proceedings of the 36th Annual ACM Symposium on User Interface Software and Technology, 2023. doi: 10.1145/3586183.3606763. URLhttps://arxiv.org/abs/2304.03442. 17

work page doi:10.1145/3586183.3606763 2023
[14]

Optimistic replication.ACM Computing Surveys, 37(1):42–81, 2005

Yasushi Saito and Marc Shapiro. Optimistic replication.ACM Computing Surveys, 37(1):42–81, 2005. doi: 10.1145/1057977.1057980. URLhttps://doi.org/10.1145/1057977.1057980

work page doi:10.1145/1057977.1057980 2005
[15]

Toolformer: Language Models Can Teach Themselves to Use Tools

Timo Schick, Jane Dwivedi-Yu, Roberto Dess`ı, Roberta Raileanu, Maria Lomeli, Eric Hambro, Luke Zettlemoyer, Nicola Cancedda, and Thomas Scialom. Toolformer: Language models can teach them- selves to use tools.arXiv preprint arXiv:2302.04761, 2023. URL https://arxiv.org/abs/2302. 04761

work page internal anchor Pith review Pith/arXiv arXiv 2023
[16]

A comprehensive study of convergent and commutative replicated data types

Marc Shapiro, Nuno Pregui c ¸a, Carlos Baquero, and Marek Zawirski. A comprehensive study of convergent and commutative replicated data types. Technical Report RR-7506, INRIA, 2011. URL https://pages.lip6.fr/Marek.Zawirski/papers/RR-7506.pdf

work page 2011
[18]

URLhttps://arxiv.org/abs/2303.11366

work page internal anchor Pith review Pith/arXiv arXiv
[20]

URLhttps://arxiv.org/abs/2506.21605

work page internal anchor Pith review Pith/arXiv arXiv
[21]

Terry, Marvin M

Douglas B. Terry, Marvin M. Theimer, Karin Petersen, Alan J. Demers, Mike J. Spreitzer, and Carl H. Hauser. Managing update conflicts in Bayou, a weakly connected replicated storage system. In Proceedings of the 15th ACM Symposium on Operating Systems Principles (SOSP). ACM, 1995. doi: 10.1145/224056.224070. URLhttps://doi.org/10.1145/224056.224070

work page doi:10.1145/224056.224070 1995
[22]

ReAct: Synergizing Reasoning and Acting in Language Models

Shunyu Yao, Jeffrey Zhao, Dian Yu, Nan Du, Izhak Shafran, Karthik Narasimhan, and Yuan Cao. ReAct: Synergizing reasoning and acting in language models.arXiv preprint arXiv:2210.03629, 2022. URL https://arxiv.org/abs/2210.03629

work page internal anchor Pith review Pith/arXiv arXiv 2022
[23]

Ori Yoran, Samuel Joseph Amouyal, Chaitanya Malaviya, Ben Bogin, Ofir Press, and Jonathan Berant. AssistantBench: Can web agents solve realistic and time-consuming tasks? InProceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, pages 8938–8968, 2024. doi: 10.18653/v1/2024.emnlp-main.505. URL https://aclanthology.org/2024.e...

work page doi:10.18653/v1/2024.emnlp-main.505 2024
[24]

A survey on the memory mechanism of LLM-based agents.ACM Transactions on Information Systems, 43(6):1–47, 2025

Zeyu Zhang, Xiaohe Bo, Chen Ma, Rui Li, Xu Chen, Quanyu Dai, Jieming Zhu, Zhenhua Dong, and Ji-Rong Wen. A survey on the memory mechanism of LLM-based agents.ACM Transactions on Information Systems, 43(6):1–47, 2025. doi: 10.1145/3748302. URL https://doi.org/10.1145/ 3748302. 18

work page doi:10.1145/3748302 2025

[1] [1]

Delta State Replicated Data Types

Paulo S´ergio Almeida, Ali Shoker, and Carlos Baquero. Delta state replicated data types.Journal of Parallel and Distributed Computing, 111:162–173, 2018. doi: 10.1016/j.jpdc.2017.08.003. URL https://arxiv.org/abs/1603.01529

work page internal anchor Pith review Pith/arXiv arXiv doi:10.1016/j.jpdc.2017.08.003 2018

[2] [2]

Evaluating Memory in LLM Agents via Incremental Multi-Turn Interactions

Yuanzhe Hu, Yu Wang, and Julian McAuley. Evaluating memory in LLM agents via incremental multi-turn interactions.arXiv preprint arXiv:2507.05257, 2025. URL https://arxiv.org/abs/ 2507.05257. Introduces MemoryAgentBench

work page internal anchor Pith review Pith/arXiv arXiv 2025

[3] [3]

AMC: Towards trustworthy and explorable CRDT applications with the Automerge model checker

Andrew Jeffery and Richard Mortier. AMC: Towards trustworthy and explorable CRDT applications with the Automerge model checker. InWorkshop on Principles and Practice of Consistency for Distributed Data (PaPoC). ACM, 2023. doi: 10.1145/3578358.3591326. URL https://doi.org/ 10.1145/3578358.3591326

work page doi:10.1145/3578358.3591326 2023

[4] [4]

Beresford

Martin Kleppmann and Alastair R. Beresford. A conflict-free replicated JSON datatype.IEEE Transac- tions on Parallel and Distributed Systems, 28(10):2733–2746, 2017. doi: 10.1109/TPDS.2017.2697382. URLhttps://doi.org/10.1109/TPDS.2017.2697382

work page doi:10.1109/tpds.2017.2697382 2017

[5] [6]

URLhttps://arxiv.org/abs/1805.04263

work page internal anchor Pith review Pith/arXiv arXiv

[6] [7]

Local-first software: You own your data, in spite of the cloud

Martin Kleppmann, Adam Wiggins, Peter van Hardenberg, and Mark McGranaghan. Local-first software: You own your data, in spite of the cloud. InProceedings of the 2019 ACM SIGPLAN International Symposium on New Ideas, New Paradigms, and Reflections on Programming and Software, pages 154–178, 2019. doi: 10.1145/3359591.3359737. URL https://martin.kleppmann.c...

work page doi:10.1145/3359591.3359737 2019

[7] [8]

Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks

Patrick Lewis, Ethan Perez, Aleksandra Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich K¨uttler, Mike Lewis, Wen-tau Yih, Tim Rockt¨aschel, Sebastian Riedel, and Douwe Kiela. Retrieval-augmented generation for knowledge-intensive NLP tasks.Advances in Neural Information Processing Systems, 33:9459–9474, 2020. URLhttps://arxiv.org/abs/2005.11401

work page internal anchor Pith review Pith/arXiv arXiv 2020

[8] [9]

AgentGit: A version control framework for reliable and scalable LLM-powered multi-agent systems.arXiv preprint arXiv:2511.00628, 2025

Yang Li, Siqi Ping, Xiyu Chen, Xiaojian Qi, Zigan Wang, Ye Luo, and Xiaowei Zhang. AgentGit: A version control framework for reliable and scalable LLM-powered multi-agent systems.arXiv preprint arXiv:2511.00628, 2025. URLhttps://arxiv.org/abs/2511.00628

work page arXiv 2025

[9] [10]

GAIA: a benchmark for General AI Assistants

Gr´egoire Mialon, Cl´ementine Fourrier, Craig Swift, Thomas Wolf, Yann LeCun, and Thomas Scialom. GAIA: A benchmark for general AI assistants. InICLR 2024, 2024. URL https://arxiv.org/ abs/2311.12983

work page internal anchor Pith review Pith/arXiv arXiv 2024

[10] [11]

PROV-DM: The PROV data model

Luc Moreau and Paolo Missier. PROV-DM: The PROV data model. W3C Recommendation, 2013. URLhttps://www.w3.org/TR/prov-dm/

work page 2013

[11] [12]

MemGPT: Towards LLMs as Operating Systems

Charles Packer, Vivian Fang, Shishir Patil, Kevin Lin, Joseph E. Gonzalez, and Christopher R ´e. MemGPT: Towards LLMs as operating systems.arXiv preprint arXiv:2310.08560, 2023. URL https://arxiv.org/abs/2310.08560

work page internal anchor Pith review Pith/arXiv arXiv 2023

[12] [13]

Bernstein

Joon Sung Park, Joseph O’Brien, Carrie J. Cai, Meredith Ringel Morris, Percy Liang, and Michael S. Bernstein. Generative agents: Interactive simulacra of human behavior.Proceedings of the 36th Annual ACM Symposium on User Interface Software and Technology, 2023. doi: 10.1145/3586183.3606763. URLhttps://arxiv.org/abs/2304.03442. 17

work page doi:10.1145/3586183.3606763 2023

[13] [14]

Optimistic replication.ACM Computing Surveys, 37(1):42–81, 2005

Yasushi Saito and Marc Shapiro. Optimistic replication.ACM Computing Surveys, 37(1):42–81, 2005. doi: 10.1145/1057977.1057980. URLhttps://doi.org/10.1145/1057977.1057980

work page doi:10.1145/1057977.1057980 2005

[14] [15]

Toolformer: Language Models Can Teach Themselves to Use Tools

Timo Schick, Jane Dwivedi-Yu, Roberto Dess`ı, Roberta Raileanu, Maria Lomeli, Eric Hambro, Luke Zettlemoyer, Nicola Cancedda, and Thomas Scialom. Toolformer: Language models can teach them- selves to use tools.arXiv preprint arXiv:2302.04761, 2023. URL https://arxiv.org/abs/2302. 04761

work page internal anchor Pith review Pith/arXiv arXiv 2023

[15] [16]

A comprehensive study of convergent and commutative replicated data types

Marc Shapiro, Nuno Pregui c ¸a, Carlos Baquero, and Marek Zawirski. A comprehensive study of convergent and commutative replicated data types. Technical Report RR-7506, INRIA, 2011. URL https://pages.lip6.fr/Marek.Zawirski/papers/RR-7506.pdf

work page 2011

[16] [18]

URLhttps://arxiv.org/abs/2303.11366

work page internal anchor Pith review Pith/arXiv arXiv

[17] [20]

URLhttps://arxiv.org/abs/2506.21605

work page internal anchor Pith review Pith/arXiv arXiv

[18] [21]

Terry, Marvin M

Douglas B. Terry, Marvin M. Theimer, Karin Petersen, Alan J. Demers, Mike J. Spreitzer, and Carl H. Hauser. Managing update conflicts in Bayou, a weakly connected replicated storage system. In Proceedings of the 15th ACM Symposium on Operating Systems Principles (SOSP). ACM, 1995. doi: 10.1145/224056.224070. URLhttps://doi.org/10.1145/224056.224070

work page doi:10.1145/224056.224070 1995

[19] [22]

ReAct: Synergizing Reasoning and Acting in Language Models

Shunyu Yao, Jeffrey Zhao, Dian Yu, Nan Du, Izhak Shafran, Karthik Narasimhan, and Yuan Cao. ReAct: Synergizing reasoning and acting in language models.arXiv preprint arXiv:2210.03629, 2022. URL https://arxiv.org/abs/2210.03629

work page internal anchor Pith review Pith/arXiv arXiv 2022

[20] [23]

Ori Yoran, Samuel Joseph Amouyal, Chaitanya Malaviya, Ben Bogin, Ofir Press, and Jonathan Berant. AssistantBench: Can web agents solve realistic and time-consuming tasks? InProceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, pages 8938–8968, 2024. doi: 10.18653/v1/2024.emnlp-main.505. URL https://aclanthology.org/2024.e...

work page doi:10.18653/v1/2024.emnlp-main.505 2024

[21] [24]

A survey on the memory mechanism of LLM-based agents.ACM Transactions on Information Systems, 43(6):1–47, 2025

Zeyu Zhang, Xiaohe Bo, Chen Ma, Rui Li, Xu Chen, Quanyu Dai, Jieming Zhu, Zhenhua Dong, and Ji-Rong Wen. A survey on the memory mechanism of LLM-based agents.ACM Transactions on Information Systems, 43(6):1–47, 2025. doi: 10.1145/3748302. URL https://doi.org/10.1145/ 3748302. 18

work page doi:10.1145/3748302 2025