Pith. sign in

REVIEW 4 major objections 4 minor 50 references

This paper argues that personal-AI memory should be redesigned as a governed lifecycle, kept auditable by a shared contract of typed evidence, diagnostic traces, strategy artifacts, and rollback records.

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 · deepseek-v4-flash

2026-08-01 13:47 UTC pith:VINVBEIC

load-bearing objection A genuinely useful organizing framework with unusually honest claim boundaries, but the central cross-module audit-contract claim is openly unvalidated and needs a joint ablation before it can be taken as demonstrated. the 4 major comments →

arxiv 2607.18975 v1 pith:VINVBEIC submitted 2026-07-21 cs.AI

Mi-Memory: A Lifecycle Memory Framework for Personal AI

Xule Liu , Hanlin Teng , Chao Li , Yanan Ni , Shuo Lu , Audrey Wang , Yijun Liu , Yunfei Wang
show 10 more authors
Xiaofeng Li Xian Yi Yuanfa Li Kang Zhao Jian Liang Yuxuan Chen Jinyuan Chen Heng Qu Kun Shao Jian Luan
This is my paper
classification cs.AI
keywords personal AImemory lifecycleaudit contractevidence provenancememory governancemultimodal memoryedge deploymentstrategy evolution
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.

The report argues that long-term memory for personal assistants—across phones, cars, homes, and wearables—should be built as an auditable lifecycle, not a passive store of past chats. It proposes a shared 'lifecycle audit contract' that ties together four roles: structuring durable state, expanding evidence beyond dialogue, governing policy changes, and deploying under edge constraints. Four artifact families carry the contract: typed evidence payloads with provenance, diagnostic traces that localize where evidence is lost, versioned strategy artifacts, and gate/rollback records. Concrete modules instantiate each role, and the report reports controlled-reference accuracy on three memory benchmarks, module-level results on multimodal and cross-device evidence, a gated strategy-evolution gain on LoCoMo, and transfer-feasibility on a repository-native lightweight substrate. The paper is careful to bound each claim's evidence level and explicitly notes that cross-module joint ablation is still missing.

Core claim

The central discovery is that the 'lifecycle audit contract' can serve as the integration primitive for Personal AI memory. Rather than optimizing a single memory store, the framework defines three operational requirements for a valid memory system: evidence preservation (every answer traces back to source observations), stage-local diagnosis (a failure trace identifies the earliest loss point in ingestion, storage, retrieval, filtering, packing, or generation), and auditable evolution (every policy change carries a versioned strategy diff, a fixed harness, a gate decision, and a rollback point). The contract is carried across modules by four artifact families—typed evidence payloads, diagno

What carries the argument

The central object is the lifecycle audit contract: a set of four artifact families (typed evidence payloads that preserve source identity and provenance; diagnostic traces that record where evidence moved or disappeared; versioned strategy artifacts that make memory-policy changes explicit; and gate/rollback records that bound accepted evolution). It works by requiring every memory-dependent answer to be traceable back through context and evidence, and every policy update to carry a versioned diff, fixed harness, gate decision, and rollback point. This contract is what links the four roles—Structure, Expansion, Evolution, Deployment—and what makes failures localizable and changes reversible

Load-bearing premise

The load-bearing premise is that the shared audit contract actually improves cross-module auditability, diagnosability, and regression-resistance in a closed-loop setting; the paper's own Section 8.2 acknowledges this premise is untested ('marginal contribution still requires a joint ablation benchmark').

What would settle it

A concrete falsifier: run a joint ablation where the audit contract artifacts (typed payloads, diagnostic traces, strategy artifacts, gate/rollback records) are removed one at a time from an integrated Mi-Memory pipeline, and measure whether failure localization accuracy, silent-regression frequency, or rollback success degrade. If removing the contract artifacts does not worsen any of these metrics (or if the traces cannot predict the true failing stage better than chance), the central claim of the framework is unsupported.

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

If this is right

  • If the audit contract is adopted, memory-system failures can be attributed to a specific stage (ingestion, retrieval, filtering, packing, generation) rather than hidden in a black-box RAG pipeline.
  • Memory-policy changes become versioned, gated, and reversible, reducing silent regression risk during iteration.
  • The same contract can transfer across deployment substrates: a repository-native Markdown/Git implementation retained 90% of the service-side improvement over a no-memory baseline in a LoCoMo-aligned setting.
  • Non-dialogue evidence (images, device events, cross-device links) can enter memory as provenance-preserving payloads instead of being discarded or mis-fused.
  • The framework separates serving correctness from update correctness, making 'was this answer grounded?' and 'is this policy change safe?' distinct, auditable questions.

Where Pith is reading between the lines

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

  • If the contract holds, the next natural target is a standardized cross-system memory interface—analogous to tool-use or function-calling schemas—so different memory architectures can be jointly ablated and composed; the paper itself lists this as an open direction.
  • The paper's own Section 8.2 admits that cross-module compatibility observations are 'not a substitute for joint ablation'; until such an ablation shows that the contract improves failure localization or prevents regressions in a closed loop, the framework's main added value remains an untested hypothesis.
  • A testable extension: instrument the diagnostic traces to quantify how often the earliest-loss stage matches the true root cause (relative to oracle labels) in a joint pipeline; the paper reports stage labels but not a measure of attribution accuracy.
  • The E2MEND gain (75.58%→94.74% on LoCoMo) is descriptive, not statistical; a repeated-run study with paired significance tests and category-level confidence intervals would substantiate or weaken the claim that gated strategy search beats manual iteration.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 4 minor

Summary. The paper proposes Mi-Memory, a lifecycle memory framework for personal AI organized around four roles: Structure (MemStack), Expansion (MemSense/MemFuse), Evolution (D2ACCI/E2MEND), and Deployment (LiteMem). Its central proposal is a "lifecycle audit contract" carried by four artifact families: typed evidence payloads, diagnostic traces, strategy artifacts, and gate/rollback records. The paper argues that this contract makes evidence movement traceable across module boundaries, localizes failures to pipeline stages, and bounds policy evolution. Empirical evidence is role-specific and explicitly calibrated: MemStack is evaluated as a controlled reference on LoCoMo, PersonaMem-V2, and LongMemEval; MemSense is module-level on Mem-Gallery; MemFuse is preliminary/internal on MemFuseBench; E2MEND is a descriptive offline strategy-search result on LoCoMo; LiteMem is transfer-feasibility evidence in a repository-native setting. The paper is unusually careful in labeling these evidence boundaries, and Section 8 centralizes a claim ledger with explicit limitations.

Significance. If the lifecycle audit contract holds, it would give the community a useful organizing vocabulary for memory-system research, separating serving correctness from update correctness and making memory-policy changes inspectable and reversible. The paper's strengths include: an explicit evidence-level taxonomy, controlled-reference comparisons where possible, artifact-level reproducibility claims for strategy evolution, and unusually candid limitation statements. The formulation itself is valuable even before a full end-to-end validation. However, the main integrative claim—that the shared audit contract improves cross-module audibility, failure localization, and regression resistance—is not yet supported by evidence; it is supported by design and interface compatibility. The paper's own limitation statement in Section 8.2 acknowledges this gap. Because the framework is proposed as a systems contribution, the missing joint evaluation is the key obstacle to acceptance.

major comments (4)
  1. [§8.2] The central claim that the lifecycle audit contract integrates the four roles is left unevaluated. The paper states explicitly: "These compatibility observations are not a substitute for joint ablation" and "marginal contribution still requires a joint ablation benchmark." No experiment feeds MemFuse payloads into MemStack, evolves a strategy under E2MEND, and checks whether the shared artifacts localize failures or enable rollback in a closed loop. Since the audit contract is the paper's primary contribution (§1, §3.1), this is a load-bearing missing evaluation. A concrete remedy would be a joint ablation with and without the shared artifact families, measuring failure-localization accuracy and recovery from induced regressions.
  2. [§6.3, Table 10] The E2MEND improvement from 75.58% to 94.74% on LoCoMo (+19.16pp) is a single descriptive run in which strategy search, gate decisions, and checkpoint selection all operate on the same benchmark. No held-out split, repeated-run variance, or confidence interval is reported; the paper itself labels the result descriptive. This makes the headline number vulnerable to benchmark overfitting and does not support a generalization claim. At minimum, the authors should evaluate on a validation split or provide per-category repeated-run statistics, and state the acceptance thresholds before reporting the result as evidence of bounded evolution.
  3. [§5.3, Tables 6–7, Figure 7] The MemSense comparison is confounded: the 89.15% Mem-Gallery score uses GPT-4.1-mini with a GPT-4o-mini judge, while the cited baselines use a unified Qwen-2.5-VL-7B backbone. The paper correctly disclaims direct numerical ranking, but the operational conclusion "IKB should be the primary evidence source" is not supported without a same-backbone ablation isolating the IKB contribution. Similarly, MemFuseBench is an internal benchmark scored by the authors' judge pipeline, with no human upper bound or inter-annotator agreement; the +4.7pp margin over mem0 is preliminary and should not be read as cross-device fusion superiority beyond descriptive evidence.
  4. [§3.1, Appendix E, Eq. (14)] The audit contract's second requirement, stage-local diagnosis, is never independently validated. Diagnostic traces are emitted by the same pipeline that is being audited, and the evidence-preservation ladder in Eq. (14) assumes knowledge of ground-truth evidence. No experiment measures trace-label accuracy against human-annotated failure causes or independent instrumentation. Without such validation, the core diagnostic property of the contract is assumed rather than demonstrated. A trace-label accuracy study, even on a small sample, would materially strengthen the claim.
minor comments (4)
  1. [§4.3 vs §8.1] The LoCoMo margin over MemBrain is reported as +0.36% in Figure 4 and +0.34pp in Section 8.1. These should be reconciled or explicitly described as different quantities (percentage vs. percentage-point change).
  2. [§5.1, Eq. (5)] The provenance-relevance trade-off weight λ is introduced without any value or sensitivity analysis. Since the objective is used to justify the Evidence Admission contract, a brief statement of the chosen value and its stability would help.
  3. [§7.3] The LiteMem transfer evidence uses a LoCoMo-aligned setting with a modular TypeScript harness. The paper should state explicitly that the 90.81% score is not directly comparable to the MemStack LoCoMo result because of differences in harness, retrieval surface, and evaluation prompts, even though the retention calculation is transparent.
  4. [Appendix G, Eq. (20)] The LiteMem scoring equation uses four λ weights and two decay timescales, but the paper does not report the configured values or any sensitivity analysis. These are implementation-level parameters, but a brief configuration note would improve reproducibility.

Circularity Check

1 steps flagged

E2MEND's LoCoMo gain is a selected optimum on the same benchmark; otherwise the paper is transparent and the audit contract is a definition rather than a derivation.

specific steps
  1. fitted input called prediction [Section 6.3 (E2MEND validation) and Algorithm 2 (Appendix F)]
    "Run full paired item-level comparison between c and s_t ... if configured acceptance and bounded regression checks pass ... On LoCoMo, E2MEND improves the initial memory-system baseline from75.58%(1164/1540) to the reported best checkpoint of94.74%(1459/1540), a +19.16pp gain and +295 net correct answers."

    E2MEND's search and acceptance use the same LoCoMo benchmark as both the optimization objective and the reported validation. Algorithm 2 accepts a candidate only after a "full paired item-level comparison" on that fixed benchmark, so the reported best checkpoint is the maximum over strategies selected by the gate on the same data. The +19.16pp is therefore the optimized value of the selection criterion, not an independent out-of-sample result. The paper is transparent ("descriptive full-benchmark evidence", "not a repeated-run statistical claim"), but the improvement is forced by the acceptance rule rather than being a prediction that could fail.

full rationale

The paper's central contribution is a normative audit contract, not a first-principles derivation, so most of the framework is definitional rather than circular. MemStack is anchored to external benchmarks (LoCoMo, PersonaMem-V2, LongMemEval) with a reproduced MemBrain baseline; MemSense uses the public Mem-Gallery benchmark; LiteMem's 90.0% retention is arithmetic on reported numbers. The only reduction-by-construction I find is the E2MEND LoCoMo gain, because the same benchmark is both the search objective and the validation surface; the paper explicitly labels this "descriptive full-benchmark evidence" and "not a repeated-run statistical claim." Section 8.2 also concedes "These compatibility observations are not a substitute for joint ablation" and "marginal contribution still requires a joint ablation benchmark," which is a load-bearing missing evaluation rather than an internal circularity. Self-citations (HarnessX [5], MIA [27]) are contextual and not load-bearing. Overall: partial circularity in one evidence anchor; the central framework retains independent content and is transparent about its boundaries.

Axiom & Free-Parameter Ledger

6 free parameters · 6 axioms · 6 invented entities

The framework's claims rest largely on internal representations, self-defined artifacts, and benchmarks that are either external but single-run or internal with no independent validation. The main free parameters are retrieval/evolution/lite-memory tuning constants, most of which are hand-set or searched on the evaluation benchmark itself. The audit contract, the layered memory, IKB, and fusion constructs are invented design entities with only internal evidence.

free parameters (6)
  • RRF channel weights and κ = Uniform baseline; κ=60 unless adapter overrides
    Retrieval fusion parameters in Eq. 10 are set by convention and can be evolved by E2MEND; not derived from first principles.
  • Context budget reservations = L2 slice typically 10–20%; forget/update constraints budget-exempt
    Table 20 default packing policy; affects which evidence reaches generation and is set by design/tuning.
  • Dedup threshold = 0.92 after t3→t4 iteration (0.75 reverted)
    Selected on LoCoMo dev traces in D2ACCI round t4 to fix ingestion-gap failures; a fitted decision threshold, not a prediction.
  • E2MEND strategy-space values = Varied across stages S0–S5; final best checkpoint 94.74% on LoCoMo
    Mutable fields in Table 26 (top_k, rerank top_n, BM25/vector weights, decay half-life, prompts) are searched on the benchmark; the +19.16pp gain is a fitting result.
  • LiteMem ranking coefficients = Configuration-level values fixed within simulated LoCoMo timeline
    Eqs. 20–21 involve λ1–λ4, τ, τm, η, ρ; sensitivity is not reported, and values are chosen for the transfer run.
  • MemFuse λ (Eq. 5) = Implementation-specific; no value reported
    Balances relevance vs provenance in evidence admission; no fitting details provided.
axioms (6)
  • domain assumption LoCoMo, PersonaMem-V2, LongMemEval, and Mem-Gallery are valid measures of memory continuity and grounding.
    Used as ground truth in Sections 4.3 and 5.3; no verification that benchmark accuracy tracks real auditable memory behavior.
  • domain assumption The reproduced MemBrain baseline and EverMemOS evaluation harness are faithful controlled references.
    MemBrain and EverMemOS are not cited or released; Figure 4's comparative margins rest on this unstated reproduction.
  • domain assumption Complete evidence annotations are available for Layer-A diagnosis on LoCoMo.
    Algorithm 3 in Appendix F.4 depends on complete evidence ids; the paper says in deployment-like settings Layer A is heuristic and audit-assisted.
  • domain assumption LLM judges used in evaluations are reliable.
    No inter-annotator agreement or human upper bound is reported for MemFuseBench; gpt-4o-mini/gpt-4o judges are used.
  • domain assumption Memory behavior can be evaluated within a trusted serving environment with platform-level privacy and safety handled separately.
    Section 3.1 explicitly scopes out privacy enforcement, consent, and abuse handling; the audit contract's sufficiency depends on that boundary.
  • domain assumption Threshold-based gates and rollback rules are sufficient to prevent silent regressions.
    No formal proof or closed-loop experiment; E2MEND relies on hard gates, a Critic, and one descriptive LoCoMo run.
invented entities (6)
  • Lifecycle audit contract no independent evidence
    purpose: Unifying organizing principle for typed evidence, diagnostic traces, strategy artifacts, and gate/rollback records across roles.
    Core conceptual contribution; no external falsifiable test of its benefit; paper says joint ablation is still needed (Section 8.2).
  • L0/L1/L2/SM layered memory no independent evidence
    purpose: MemStack's four-layer decomposition for observability and continuity.
    Internal representation; benchmark results support usefulness but the layer taxonomy itself has no independent validation.
  • Image Knowledge Base (IKB) no independent evidence
    purpose: MemSense's structured side table for visual identity, routing, and provenance.
    Evaluated on Mem-Gallery at module level, but no controlled ablation isolates the IKB contribution (Section 5.3, Table 23).
  • MemoryPack / FusedNode / FusionSession no independent evidence
    purpose: MemFuse cross-device causal fusion constructs.
    Evaluated only on internal MemFuseBench; external validation is pending (Section 5.3).
  • ProcedureEntry (procedural hooks) no independent evidence
    purpose: Store user-specific conversational response procedures in MemStack.
    Design-only; explicitly not benchmarked (Sections 4.2, B.1).
  • Strategy artifact / gate / rollback records no independent evidence
    purpose: E2MEND's versioned strategy and governance artifacts.
    Internal to the framework; the supporting evidence is a descriptive LoCoMo run with no repeated-run statistics.

pith-pipeline@v1.3.0-alltime-deepseek · 41533 in / 15696 out tokens · 148104 ms · 2026-08-01T13:47:43.169381+00:00 · methodology

0 comments
read the original abstract

Personal AI is moving beyond chat-only interaction toward continuous services that span phones, cars, homes, wearables, cameras, and tools. In this setting, memory cannot remain a cache of prior conversations. It should serve as a continuity and governance substrate: preserving durable user state, grounding answers in multimodal and device evidence, supporting correction and forgetting, bounding policy evolution, and remaining deployable under latency, cost, privacy, and edge-cloud constraints. This technical report presents Mi-Memory, a lifecycle memory framework for Personal AI organized around four roles: Structure, Expansion, Evolution, and Deployment. A shared audit contract links these roles through four recurring artifact families: typed evidence payloads preserve source identity and provenance, diagnostic traces localize evidence loss across the serving pipeline, strategy artifacts make memory-policy changes explicit, and gate/rollback records bound accepted evolution. MiMemory instantiates the roles through MemStack, MemSense/MemFuse, D$^{2}$ACCI/E$^{2}$MEND, and LiteMem. In controlled-reference Structure evaluations, MemStack reaches 93.59%, 57.24%, and 87.47% on LoCoMo, PersonaMem-V2, and LongMemEval, respectively; other tracks report module-level, preliminary/internal, transfer-feasibility, or design-only evidence with explicit boundaries. MiMemory is a step toward auditable, evidence-gated, and deployment-aware memory systems for Personal AI. Project homepage: https://darwin-agent.github.io/Mi-Memory/ .

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

50 extracted references · 35 linked inside Pith

  1. [1]

    Mem-Gallery: Benchmarking multimodal long-term conversational memory for MLLM agents.arXiv preprint arXiv:2601.03515, 2026

    Yuanchen Bei, Tianxin Wei, Xuying Ning, Yanjun Zhao, Zhining Liu, Xiao Lin, Yada Zhu, Hendrik Hamann, Jingrui He, and Hanghang Tong. Mem-Gallery: Benchmarking multimodal long-term conversational memory for MLLM agents.arXiv preprint arXiv:2601.03515, 2026

    arXiv 2026

  2. [2]

    Varun Pratap Bhardwaj. SuperLocalMemory v3.3: The living brain—biologically-inspired forgetting, cognitive quantization, and multi-channel retrieval for zero-LLM agent memory systems.arXiv preprint arXiv:2604.04514, 2026

    Pith/arXiv arXiv 2026

  3. [3]

    Choquette-Choo, Hengrui Jia, Adelin Travers, Baiwu Zhang, David Lie, and Nicolas Papernot

    Lucas Bourtoule, Varun Chandrasekaran, Christopher A. Choquette-Choo, Hengrui Jia, Adelin Travers, Baiwu Zhang, David Lie, and Nicolas Papernot. Machine unlearning. InIEEE Symposium on Security and Privacy, pages 141–159, 2021

    2021

  4. [4]

    STALE: Can LLM agents know when their memories are no longer valid?arXiv preprint arXiv:2605.06527, 2026

    Hanxiang Chao, Yihan Bai, Rui Sheng, Tianle Li, and Yushi Sun. STALE: Can LLM agents know when their memories are no longer valid?arXiv preprint arXiv:2605.06527, 2026

    Pith/arXiv arXiv 2026

  5. [5]

    HarnessX: A composable, adaptive, and evolvable agent harness foundry.arXiv preprint arXiv:2606.14249, 2026

    Tingyang Chen, Shuo Lu, Kang Zhao, Weicheng Meng, Hanlin Teng, Tianhao Li, Chao Li, Xule Liu, Jian Liang, Zhizhong Zhang, Yuan Xie, Heng Qu, Kun Shao, and Jian Luan. HarnessX: A composable, adaptive, and evolvable agent harness foundry.arXiv preprint arXiv:2606.14249, 2026

    Pith/arXiv arXiv 2026

  6. [6]

    Exploring cross-scenario generality of agentic memory systems: Diagnostics and a strong baseline

    Zhikai Chen, Jialiang Gu, Junyu Yin, Xianxuan Long, Shenglai Zeng, Xiaoze Liu, Kai Guo, Keren Zhou, and Jiliang Tang. Exploring cross-scenario generality of agentic memory systems: Diagnostics and a strong baseline. arXiv preprint arXiv:2606.04315, 2026

    Pith/arXiv arXiv 2026

  7. [7]

    Mem0: Building production- ready AI agents with scalable long-term memory.arXiv preprint arXiv:2504.19413, 2025

    Prateek Chhikara, Dev Khant, Saket Aryan, Taranjeet Singh, and Deshraj Yadav. Mem0: Building production- ready AI agents with scalable long-term memory.arXiv preprint arXiv:2504.19413, 2025

    Pith/arXiv arXiv 2025

  8. [8]

    When generic prompt improvements hurt: Evaluation-driven iteration for LLM applications

    Daniel Commey. When generic prompt improvements hurt: Evaluation-driven iteration for LLM applications. arXiv preprint arXiv:2601.22025, 2026

    Pith/arXiv arXiv 2026

  9. [9]

    LightMem: Lightweight and efficient memory-augmented generation.arXiv preprint arXiv:2510.18866, 2025

    Jizhan Fang, Xinle Deng, Haoming Xu, Ziyan Jiang, Yuqi Tang, Ziwen Xu, Shumin Deng, Yunzhi Yao, Mengru Wang, Shuofei Qiao, Huajun Chen, and Ningyu Zhang. LightMem: Lightweight and efficient memory-augmented generation.arXiv preprint arXiv:2510.18866, 2025

    Pith/arXiv arXiv 2025

  10. [10]

    Memory beyond recall: A dual-process cognitive memory system for self-evolving LLM agents.arXiv preprint arXiv:2606.09483, 2026

    Tianxiang Fei, Mingyang Song, Mao Zheng, and Xiang Yu. Memory beyond recall: A dual-process cognitive memory system for self-evolving LLM agents.arXiv preprint arXiv:2606.09483, 2026

    Pith/arXiv arXiv 2026

  11. [11]

    M2A: Multimodal memory agent with dual-layer hybrid memory for long-term personalized interactions.arXiv preprint arXiv:2602.07624, 2026

    Junyu Feng, Binxiao Xu, Jiayi Chen, Mengyu Dai, Cenyang Wu, Haodong Li, Bohan Zeng, Yunliu Xie, Hao Liang, Ming Lu, and Wentao Zhang. M2A: Multimodal memory agent with dual-layer hybrid memory for long-term personalized interactions.arXiv preprint arXiv:2602.07624, 2026

    arXiv 2026

  12. [12]

    MemoryArena: Benchmarking agent memory in interdependent multi-session agentic tasks.arXiv preprint arXiv:2602.16313, 2026

    Zexue He, Yu Wang, Churan Zhi, Yuanzhe Hu, Tzu-Ping Chen, Lang Yin, Ze Chen, Tong Arthur Wu, Siru Ouyang, Zihan Wang, Jiaxin Pei, Julian McAuley, Yejin Choi, and Alex Pentland. MemoryArena: Benchmarking agent memory in interdependent multi-session agentic tasks.arXiv preprint arXiv:2602.16313, 2026

    arXiv 2026

  13. [13]

    AMA: Adaptive memory via multi-agent collaboration.arXiv preprint arXiv:2601.20352, 2026

    Weiquan Huang, Zixuan Wang, Hehai Lin, Sudong Wang, Bo Xu, Qian Li, Beier Zhu, Linyi Yang, and Chengwei Qin. AMA: Adaptive memory via multi-agent collaboration.arXiv preprint arXiv:2601.20352, 2026

    Pith/arXiv arXiv 2026

  14. [14]

    M 3Exam: Benchmarking multimodal memory for realistic user-agent interactions

    Zhengjun Huang, Wenxuan Liu, Zhoujin Tian, Wei Chen, Junle Chen, Yuqian Wu, Fangyuan Zhang, Qintian Guo, and Xiaofang Zhou. M 3Exam: Benchmarking multimodal memory for realistic user-agent interactions. arXiv preprint arXiv:2606.07402, 2026

    Pith/arXiv arXiv 2026

  15. [15]

    PersonaMem-V2: Towards personalized intelligence via learning implicit user personas and agentic memory.arXiv preprint arXiv:2512.06688, 2025

    Bowen Jiang, Yuan Yuan, Maohao Shen, Zhuoqun Hao, Zhangchen Xu, Zichen Chen, Ziyi Liu, Anvesh Rao Vijjini, Jiashu He, Hanchao Yu, Radha Poovendran, Gregory Wornell, Lyle Ungar, Dan Roth, Sihao Chen, and Camillo Jose Taylor. PersonaMem-V2: Towards personalized intelligence via learning implicit user personas and agentic memory.arXiv preprint arXiv:2512.06688, 2025

    arXiv 2025

  16. [16]

    Governing evolving memory in LLM agents: Risks, mechanisms, and the stability and safety governed memory (SSGM) framework.arXiv preprint arXiv:2603.11768, 2026

    Chingkwun Lam, Jiaxin Li, Lingfei Zhang, and Kuo Zhao. Governing evolving memory in LLM agents: Risks, mechanisms, and the stability and safety governed memory (SSGM) framework.arXiv preprint arXiv:2603.11768, 2026

    Pith/arXiv arXiv 2026

  17. [17]

    Beyond semantic similarity: Rethinking retrieval for agentic search via direct corpus interaction.arXiv preprint arXiv:2605.05242, 2026

    Zhuofeng Li, Haoxiang Zhang, Cong Wei, Pan Lu, Ping Nie, Yi Lu, Yuyang Bai, Shangbin Feng, Hangxiao Zhu, Ming Zhong, Yuyu Zhang, Jianwen Xie, Yejin Choi, James Zou, Jiawei Han, Wenhu Chen, Jimmy Lin, Dongfu Jiang, and Yu Zhang. Beyond semantic similarity: Rethinking retrieval for agentic search via direct corpus interaction.arXiv preprint arXiv:2605.05242...

    Pith/arXiv arXiv 2026

  18. [18]

    A survey on long-term memory security in LLM agents: Attacks, defenses, and governance across the memory lifecycle

    Zehao Lin, Xixuan Hao, Renyu Fu, Shaobo Cui, Kai Chen, Chunyu Li, Zhiyu Li, and Feiyu Xiong. A survey on long-term memory security in LLM agents: Attacks, defenses, and governance across the memory lifecycle. arXiv preprint arXiv:2604.16548, 2026

    Pith/arXiv arXiv 2026

  19. [19]

    WorldMemArena: Evaluating multimodal agent memory through action-world interaction

    Chengzhi Liu, Yuzhe Yang, Sophia Xiao Pu, Yepeng Liu, Lin Long, Yichen Guo, Nuo Chen, Zhaotian Weng, Elena Kochkina, Simerjot Kaur, Charese Smiley, Xiaomo Liu, James Zou, Sheng Liu, Yuheng Bu, Songyou Peng, and Xin Eric Wang. WorldMemArena: Evaluating multimodal agent memory through action-world interaction. arXiv preprint arXiv:2605.29341, 2026

    Pith/arXiv arXiv 2026

  20. [20]

    EvolveMem: Self- evolving memory architecture via AutoResearch for LLM agents

    Jiaqi Liu, Xinyu Ye, Peng Xia, Zeyu Zheng, Cihang Xie, Mingyu Ding, and Huaxiu Yao. EvolveMem: Self- evolving memory architecture via AutoResearch for LLM agents. arXiv preprint arXiv:2605.13941, 2026

    Pith/arXiv arXiv 2026

  21. [21]

    MemVerse: Multimodal memory for lifelong learning agents.arXiv preprint arXiv:2512.03627, 2025

    Junming Liu, Yifei Sun, Weihua Cheng, Haodong Lei, Yirong Chen, Licheng Wen, Xuemeng Yang, Daocheng Fu, Pinlong Cai, Nianchen Deng, Yi Yu, Shuyue Hu, Botian Shi, and Ding Wang. MemVerse: Multimodal memory for lifelong learning agents.arXiv preprint arXiv:2512.03627, 2025. v2, updated 2026

    Pith/arXiv arXiv 2025

  22. [22]

    Hierarchical memory orchestration for personalized persistent agents.arXiv preprint arXiv:2604.01670, 2026

    Junming Liu, Yifei Sun, Weihua Cheng, Haodong Lei, Yuqi Li, Yirong Chen, and Ding Wang. Hierarchical memory orchestration for personalized persistent agents.arXiv preprint arXiv:2604.01670, 2026

    arXiv 2026

  23. [23]

    Evalu- ating very long-term conversational memory of LLM agents

    Adyasha Maharana, Dong-Ho Lee, Sergey Tulyakov, Mohit Bansal, Francesco Barbieri, and Yuwei Fang. Evalu- ating very long-term conversational memory of LLM agents. InProceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (ACL), 2024. URLhttps://arxiv.org/abs/2402.17753

    Pith/arXiv arXiv 2024

  24. [24]

    Memory as metabolism: A design for companion knowledge systems.arXiv preprint arXiv:2604.12034, 2026

    Stefan Miteski. Memory as metabolism: A design for companion knowledge systems.arXiv preprint arXiv:2604.12034, 2026

    Pith/arXiv arXiv 2026

  25. [25]

    ByteRover: Agent-native memory through LLM-curated hierarchical context.arXiv preprint arXiv:2604.01599, 2026

    Andy Nguyen, Danh Doan, Hoang Pham, Bao Ha, Dat Pham, Linh Nguyen, Hieu Nguyen, Thien Nguyen, Cuong Do, Phat Nguyen, and Toan Nguyen. ByteRover: Agent-native memory through LLM-curated hierarchical context.arXiv preprint arXiv:2604.01599, 2026

    arXiv 2026

  26. [26]

    Patil, Kevin Lin, Sarah Wooders, and Joseph E

    Charles Packer, Vivian Fang, Shishir G. Patil, Kevin Lin, Sarah Wooders, and Joseph E. Gonzalez. MemGPT: Towards LLMs as operating systems.arXiv preprint arXiv:2310.08560, 2023

    Pith/arXiv arXiv 2023

  27. [27]

    Memory intelligence agent.arXiv preprint arXiv:2604.04503, 2026

    Jingyang Qiao, Weicheng Meng, Yu Cheng, Zhihang Lin, Zhizhong Zhang, Xin Tan, Jingyu Gong, Kun Shao, and Yuan Xie. Memory intelligence agent.arXiv preprint arXiv:2604.04503, 2026. URLhttps://arxiv.org/ abs/2604.04503

    Pith/arXiv arXiv 2026

  28. [28]

    MemLens: Benchmarking multimodal long-term memory in large vision-language models.arXiv preprint arXiv:2605.14906, 2026

    Xiyu Ren, Zhaowei Wang, Yiming Du, Zhongwei Xie, Chi Liu, Xinlin Yang, Haoyue Feng, Wenjun Pan, Tianshi Zheng, Baixuan Xu, Zhengnan Li, Yangqiu Song, Ginny Wong, and Simon See. MemLens: Benchmarking multimodal long-term memory in large vision-language models.arXiv preprint arXiv:2605.14906, 2026

    Pith/arXiv arXiv 2026

  29. [29]

    GitOfThoughts: Version-controlled reasoning and agent memory you can replay, diff, and merge.arXiv preprint arXiv:2606.14470, 2026

    Pavan C Shekar, Abhishek H S, and Aswanth Krishnan. GitOfThoughts: Version-controlled reasoning and agent memory you can replay, diff, and merge.arXiv preprint arXiv:2606.14470, 2026

    Pith/arXiv arXiv 2026

  30. [30]

    When to forget: A memory governance primitive.arXiv preprint arXiv:2604.12007, 2026

    Baris Simsek. When to forget: A memory governance primitive.arXiv preprint arXiv:2604.12007, 2026

    Pith/arXiv arXiv 2026

  31. [31]

    Introducing the token-efficient memory algorithm.Mem0 Research Blog, apr 2026

    Taranjeet Singh. Introducing the token-efficient memory algorithm.Mem0 Research Blog, apr 2026. URL https://mem0.ai/blog/mem0-the-token-efficient-memory-algorithm

    2026

  32. [32]

    Beyondamilliontokens: Benchmarkingandenhancinglong-termmemoryinLLMs

    MohammadTavakolietal. Beyondamilliontokens: Benchmarkingandenhancinglong-termmemoryinLLMs. In International Conference on Learning Representations (ICLR), 2026. URLhttps://arxiv.org/abs/2510.27246

    arXiv 2026

  33. [33]

    From recall to forgetting: Benchmarking long-term memory for personalized agents.arXiv preprint arXiv:2604.20006, 2026

    Md Nayem Uddin, Kumar Shubham, Eduardo Blanco, Chitta Baral, and Gengyu Wang. From recall to forgetting: Benchmarking long-term memory for personalized agents.arXiv preprint arXiv:2604.20006, 2026

    Pith/arXiv arXiv 2026

  34. [34]

    MemMachine: A ground-truth-preserving memory system for personalized AI agents.arXiv preprint arXiv:2604.04853, 2026

    Shu Wang, Edwin Yu, Oscar Love, Tom Zhang, Tom Wong, Steve Scargall, and Charles Fan. MemMachine: A ground-truth-preserving memory system for personalized AI agents.arXiv preprint arXiv:2604.04853, 2026

    Pith/arXiv arXiv 2026

  35. [35]

    LongMemEval: Benchmark- ing chat assistants on long-term interactive memory

    Di Wu, Hongwei Wang, Wenhao Yu, Yuwei Zhang, Kai-Wei Chang, and Dong Yu. LongMemEval: Benchmark- ing chat assistants on long-term interactive memory. InInternational Conference on Learning Representations (ICLR), 2025. URLhttps://arxiv.org/abs/2410.10813

    Pith/arXiv arXiv 2025

  36. [36]

    LongMemEval- V2: Evaluating long-term agent memory toward experienced colleagues.arXiv preprint arXiv:2605.12493, 2026

    DiWu, ZixiangJi, AsmiKawatkar, BryanKwan, Jia-ChenGu, NanyunPeng, andKai-WeiChang. LongMemEval- V2: Evaluating long-term agent memory toward experienced colleagues.arXiv preprint arXiv:2605.12493, 2026

    Pith/arXiv arXiv 2026

  37. [37]

    Git context controller: Manage the context of LLM-based agents like Git.arXiv preprint arXiv:2508.00031, 2025

    Junde Wu, Minhao Hu, Jiayuan Zhu, Jiazhen Pan, Yuyuan Liu, Min Xu, and Yueming Jin. Git context controller: Manage the context of LLM-based agents like Git.arXiv preprint arXiv:2508.00031, 2025. v2, updated 2026. 32

    Pith/arXiv arXiv 2025

  38. [38]

    Memora: A harmonic memory representation balancing abstraction and specificity.arXiv preprint arXiv:2602.03315, 2026

    Menglin Xia, Xuchao Zhang, Shantanu Dixit, Paramaguru Harimurugan, Rujia Wang, Victor Ruhle, Robert Sim, Chetan Bansal, and Saravan Rajmohan. Memora: A harmonic memory representation balancing abstraction and specificity.arXiv preprint arXiv:2602.03315, 2026

    Pith/arXiv arXiv 2026

  39. [39]

    EvoArena: Tracking memory evolution for robust LLM agents in dynamic environments.arXiv preprint arXiv:2606.13681, 2026

    Jundong Xu, Qingchuan Li, Jiaying Wu, Yihuai Lan, Shuyue Stella Li, Huichi Zhou, Bowen Jiang, Lei Wang, Jun Wang, Anh Tuan Luu, Caiming Xiong, Hae Won Park, Bryan Hooi, and Zhiyuan Hu. EvoArena: Tracking memory evolution for robust LLM agents in dynamic environments.arXiv preprint arXiv:2606.13681, 2026

    Pith/arXiv arXiv 2026

  40. [40]

    Pan, Hinrich Schütze, Volker Tresp, and Yunpu Ma

    Sikuan Yan, Xiufeng Yang, Zuchao Huang, Ercong Nie, Zifeng Ding, Zonggen Li, Xiaowen Ma, Jinhe Bi, Kristian Kersting, Jeff Z. Pan, Hinrich Schütze, Volker Tresp, and Yunpu Ma. Memory-R1: Enhancing large language model agents to manage and utilize memories via reinforcement learning. arXiv preprint arXiv:2508.19828, 2025

    Pith/arXiv arXiv 2025

  41. [41]

    Data-centric artificial intelligence: A survey.ACM Computing Surveys, 57(5):1–42, 2025

    Daochen Zha, Zaid Pervaiz Bhat, Kwei-Herng Lai, Fan Yang, Zhimeng Jiang, Shaochen Zhong, and Xia Hu. Data-centric artificial intelligence: A survey.ACM Computing Surveys, 57(5):1–42, 2025. doi: 10.1145/3711118

    doi:10.1145/3711118 2025

  42. [42]

    MemSkill: Learning and evolving memory skills for self-evolving agents.arXiv preprint arXiv:2602.02474, 2026

    HaozhenZhang, QuanyuLong, JianzhuBao, TaoFeng, WeizhiZhang, HaodongYue, andWenyaWang. MemSkill: Learning and evolving memory skills for self-evolving agents.arXiv preprint arXiv:2602.02474, 2026

    Pith/arXiv arXiv 2026

  43. [43]

    Beyond similarity: Trustworthy memory search for personal AI agents.arXiv preprint arXiv:2606.06054, 2026

    Jiawen Zhang, Kejia Chen, Jiachen Ma, Yangfan Hu, Lipeng He, Yechao Zhang, Jian Liu, Xiaohu Yang, Tianwei Zhang, and Ruoxi Jia. Beyond similarity: Trustworthy memory search for personal AI agents.arXiv preprint arXiv:2606.06054, 2026

    Pith/arXiv arXiv 2026

  44. [44]

    Externalization in LLM agents: A unified review of memory, skills, protocols and harness engineering.arXiv preprint arXiv:2604.08224, 2026

    Chenyu Zhou, Huacan Chai, Wenteng Chen, Zihan Guo, Rong Shan, Yuanyi Song, Tianyi Xu, Yingxuan Yang, Aofan Yu, Weiming Zhang, Congming Zheng, Jiachen Zhu, Zeyu Zheng, Zhuosheng Zhang, Xingyu Lou, Chang- wang Zhang, Zhihui Fu, Jun Wang, Weiwen Liu, Jianghao Lin, and Weinan Zhang. Externalization in LLM agents: A unified review of memory, skills, protocols ...

    Pith/arXiv arXiv 2026

  45. [45]

    Memento: Fine-tuning LLM agents without fine-tuning LLMs.arXiv preprint arXiv:2508.16153, 2025

    Huichi Zhou, Yihang Chen, Siyuan Guo, Xue Yan, Kin Hei Lee, Zihan Wang, Ka Yiu Lee, Guchun Zhang, Kun Shao, Linyi Yang, and Jun Wang. Memento: Fine-tuning LLM agents without fine-tuning LLMs.arXiv preprint arXiv:2508.16153, 2025. URLhttps://arxiv.org/abs/2508.16153. 33 Contributions and Acknowledgments Core Contributors. •Xule Liu ∗ •Hanlin Teng ∗ •Chao L...

    Pith/arXiv arXiv 2025

  46. [46]

    Amy's colleague Rob

    Preserve all named entities verbatim -- use full identifiers (e.g. "Amy's colleague Rob"), never generic references

  47. [47]

    Retain exact quantities: numbers, prices, dates, times, percentages, frequencies

  48. [48]

    every Tuesday and Thursday

    Maintain temporal specificity -- "every Tuesday and Thursday", not "twice a week"

  49. [49]

    When multiple records describe similar events, use timestamps and layer metadata to distinguish them 38

  50. [50]

    they/we/together

    Perform inference only when evidence from multiple layers strongly supports the connection # STRUCTURED REASONING PATH ## Step 1: EVIDENCE CANDIDATES Enumerate all memory records (L0 facts, L1 summaries, L2 profile entries, SM context) that could relate to the question. Include records with unresolved relative dates -- do not filter prematurely. ## Step 2...

    2023