C3M: Cross-Session Multimodal Memory Maintenance for Long-Horizon Tasks
Xueshu Chen, Yan Wang, Zihao Xue, Jiefu Li, Zhenfang Liu, Jayden Chen, Zhen Bi, Jungang Lou
cs.AI · cs.CL · cs.CV · cs.IR
Quick verdict
C3M couples a relation-aware index with provenance-preserving budgeted routing; results lead at 32K/64K but trail CDRs at 128K and lack significance testing.
Summary
The paper proposes C3M, a memory organization for AI agents handling long-horizon, cross-session multimodal tasks under a fixed, query-blind memory budget. It separates a bounded active index of compact routing entries with provenance pointers from a persistent cold store of immutable text-image raw nodes. The RAMU maintenance module classifies each incoming session against retrieved candidates as same, update, distinct, or uncertain, then applies conservative MERGE, UPDATE, ADD, or KEEP actions, guaranteeing that merged entries retain access to all predecessor and incoming evidence (Eq. 4). At query time, BMER ranks index pages and entries, then expands selected entries into raw nodes, images, and text under a joint budget over pages, nodes, images, and tokens. Experiments use MemLens's 195-question agent subset at 32K/64K/128K history lengths with GPT-5.6 Sol and Qwen 3.8 Flash, against ReSum, MovieChat, MemRefine, and CDRs under shared store and reader budgets. C3M achieves the best overall accuracy at 32K and 64K with both models (73.33% and 68.21% with GPT-5.6 Sol) but falls behind CDRs at 128K on both. Ablations on 34 questions show index saturation as history grows and non-monotonic accuracy under budget scaling. Key limitations: no significance testing, small per-category samples (~22-61 questions), a single benchmark, and unspecified hyperparameters.
Review complete
Assessment at a glance
5
risk flags identified
● high 1
● medium 3
● low 1
5 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- Similarity-based merging is unsafe for multimodal memory; relation-aware classification (same/update/distinct/uncertain) with conservative thresholds preserves distinctions future queries may need.
- Coupling a bounded active index with a persistent cold store preserves full provenance (Eq. 4: Src(e′) ⊇ Src(e⋆) ∪ Pt) without overwriting any source evidence.
- Budgeted routing (BMER) under a joint page/node/image/text budget recovers source evidence better than ReSum, MovieChat, MemRefine, and CDRs at 32K/64K histories on MemLens.
- C3M leads Overall accuracy at 32K and 64K with both GPT-5.6 Sol and Qwen 3.8 Flash, but trails CDRs at 128K.
- Index capacity relieves structural pressure (overflow repairs 33.38→3.38) but does not monotonically improve task outcomes.
Strengths
Formalized provenance guarantee, not just a slogan
Eq. 4 requires Src(e′) ⊇ Src(e⋆) ∪ Pt; Raw Nodes are never overwritten, and updates link predecessor plus new source — a checkable invariant rather than a vague claim, verified empirically in §4.1 (pointer counts 64.79/126.97/239.32 match input nodes).
Conservative merge policy targets the right failure mode
RAMU (Eqs. 2–3) makes similarity propose candidates only; relation labels same/update/distinct/uncertain with confidence threshold δ route to MERGE/UPDATE/ADD/KEEP, with uncertainty defaulting to KEEP — directly addressing the conflation risk framed in Fig. 1.
Unusually fair shared-budget protocol
Honest reporting of losing conditions
Failure-attribution instrumentation
Weaknesses
No statistical treatment at all
Advantage does not survive 128K histories
Thin, non-monotonic ablations
Reproducibility gaps despite a code link
Single benchmark, author-defined protocol
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
C3M is a 4-page (ICASSP-style) method paper on memory for long-horizon agents in cross-session multimodal settings: maintain a bounded, query-blind active index while keeping original text-image evidence recoverable. Central claims: (1) similarity alone is unsafe for merging; relation-aware classification (same/update/distinct/uncertain, Eqs. 2–3) with a confidence threshold δ preserves necessary distinctions; (2) a persistent cold store plus provenance pointers guarantees source recovery (Eq. 4: Src(e′) ⊇ Src(e⋆) ∪ Pt); (3) budgeted routing (BMER, Eqs. 5–7) is more effective than existing compression methods under shared budgets. Claim (3) is only partially supported: Table 1 shows C3M best Overall at 32K/64K but behind CDRs at 128K on both execution models.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is C3M?
C3M is a memory organization for AI agents working on long-horizon, cross-session tasks with text and images. It keeps a small, bounded index of compact routing entries with provenance pointers, stores all original text-image evidence unmodified in a persistent store, updates the index with conservative relation-aware merge/update/add/keep decisions, and at query time retrieves original evidence under a fixed budget.
Does C3M beat all competing memory methods?
No. On the MemLens agent benchmark (195 questions), C3M has the best overall accuracy at 32K and 64K token histories with both GPT-5.6 Sol and Qwen 3.8 Flash (e.g., 73.33% and 68.21% with GPT-5.6 Sol), but the CDRs method leads at 128K histories on both models. The paper also reports no statistical significance tests, so small margins should be read cautiously.
What data and models were used to evaluate C3M?
The official 195-question agent subset of the MemLens benchmark, replayed chronologically at 32K/64K/128K history lengths across five task categories (information extraction, multi-session reasoning, temporal reasoning, knowledge update, answer refusal). GPT-5.6 Sol and Qwen 3.8 Flash built memory and answered; one shared frontier model judged answers; baselines were ReSum, MovieChat, MemRefine, and CDRs under identical memory and reader budgets.
Is code available for C3M?
The paper states code is available at github.com/HuzhouNLP/C3M. However, the paper itself does not specify key hyperparameters (merge threshold δ, candidate count, routing weights γ/λ) or the identity of the judge model, and results depend on proprietary GPT-5.6 Sol and Qwen 3.8 Flash APIs, so exact replication may require author assistance.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figures 1–4 are available only as captions/labels in the extracted text; plot values (Fig. 3 curves) and case-study imagery cannot be analyzed.
- Hyperparameter values (δ, TopK, γ, λ) and the judge model identity are absent from the text.
- Per-category sample sizes (n≈22–61) are inferred from accuracy fractions, not stated in the text.
- Extraction may be incomplete (this appears to be the full 4-page paper, but appendix-level details may be missing).
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.