CreaMem: A Scene-Aware Memory Architecture for Personalized Agents
Qixuan Sun, Yue Que, Bowei He, Jin Guo, Dihang Yang, Wenchang Situ, Chen Ma
cs.CL
Quick verdict
CreaMem partitions agent memory by life scenes and dual-codes entries; tops LoCoMo 54.61% and LongMemEval-S 66.40%, but main gains lack significance testing.
Summary
CreaMem is a memory architecture for personalized LLM agents that organizes long-term memory by the user's life scenes. A Meta Memory Manager routes each incoming message to up to three Life Scene Memories (Life, Work, Interest) while writing a paired timeline entry into a dedicated Episodic Memory; a Core Memory holds a static profile. At retrieval, a Planner LLM selects memories and keywords, and per-memory balanced sampling over BM25 and dense retrieval, fused via Reciprocal Rank Fusion, returns context. On LoCoMo, CreaMem reaches 54.61% GPT-4o-judged accuracy versus 45.62% for the strongest baseline, HippoRAG 2; on LongMemEval-S it reaches 66.40% versus 60.20% for MemGAS, at comparable token budgets. Ablations attribute most of the gain to dual episodic/trait encoding (75.1% versus 65.9% under a Gemini-3 backbone and judge), while fusion and sampling choices show statistically null effects. Gains persist under Qwen3.6 and Gemini-3 backbones, though CreaMem trails some baselines on Single-hop and Open-domain questions. Main limitations: headline comparisons lack confidence intervals and significance tests under a single LLM judge; the fixed three-scene taxonomy is validated at only 71% GPT-4o coverage; and retrieval latency rises to 2.9 seconds per query versus 1.98 for A-Mem.
Review complete
Assessment at a glance
5
risk flags identified
● high 0
● medium 2
● low 3
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
- Partitioning agent memory into Life/Work/Interest scene memories reduces cross-scene interference at retrieval and improves QA accuracy.
- Dual-coding each experience as an episodic timeline entry plus a scene-conditional trait entry accounts for the majority of CreaMem's gains, not fusion or sampling choices.
- CreaMem outperforms ten memory baselines plus full-history on LoCoMo (54.61% vs 45.62% 4o-Judge) and LongMemEval-S (66.40% vs 60.20%) at comparable token budgets.
- The gains transfer across GPT-4o-mini, Qwen3.6-35B-A3B, and Gemini-3 backbones, with the largest improvements on temporal and multi-hop questions.
Strengths
Matched-comparison protocol
All Table 1 methods share the GPT-4o-mini backbone, MemGAS QA/judge prompts, and a GPT-4o judge; CreaMem's retrieval budget is deliberately held at k=20, below its k=30 accuracy optimum (Fig. 2), to match baseline token budgets (2,805 avg tokens on LoCoMo vs 1,931-3,074 for structured baselines).
Honest null-result reporting in the appendix
Appendix Table 4 shows RRF changes accuracy by +0.06/-0.91 points (both 95% CIs span zero) and balanced-vs-dynamic sampling by +0.26 points (CI [-0.78, 1.30]), transparently locating the gain in dual encoding rather than the retrieval machinery.
Design axes cleanly isolated
System reliability audits
Cross-backbone validation
Weaknesses
No uncertainty for headline numbers
Baseline number inconsistency
Aggregate 'all metrics' framing vs category losses
Retrieval machinery unvalidated as contributors
Taxonomy coverage stretched
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
CreaMem (arXiv 2609.08550, cs.CL) proposes a scene-aware memory architecture for personalized LLM agents. The authors identify two gaps in prior memory systems - no organization by the user's life scenes, and single-perspective encoding of each experience - and address both: memory is partitioned into Life, Work, and Interest scene memories plus a dedicated Episodic Memory and a Core Memory (Sec. 3.1-3.2); every experience is dual-coded as a timeline entry (Eq. 1) and a scene-conditional trait entry (Eq. 3); retrieval combines per-memory balanced BM25+dense candidates via Reciprocal Rank Fusion (Sec. 3.4). Headline claims: best QA accuracy on LoCoMo and LongMemEval-S, largest gains on multi-hop and temporal questions, and gains driven by organization rather than context size.
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 CreaMem and what problem does it solve?
CreaMem is a memory architecture for personalized LLM agents. It splits long-term memory into three life-scene stores (Life, Work, Interest) plus an Episodic Memory for dated events and a Core Memory for the user profile, and stores every experience twice - as a timeline event and as a scene-specific trait - so retrieval can combine both views while memories from unrelated life scenes do not compete for the same retrieval slots.
How much does CreaMem improve accuracy over prior memory systems?
With GPT-4o-mini as the backbone and GPT-4o as judge, CreaMem scores 54.61% on LoCoMo versus 45.62% for the strongest prior baseline (HippoRAG 2), and 66.40% on LongMemEval-S versus 60.20% for MemGAS, while using a comparable number of retrieved tokens (about 2,805 on LoCoMo). The largest relative gains are on temporal and multi-hop questions.
Is CreaMem's code and data available for reproduction?
The authors state that code is released in a public GitHub repository, and both benchmarks (LoCoMo and LongMemEval-S) are public. Key prompts and hyperparameters (temperature 0, RRF smoothing k0=60, top-k 20/40, batch extraction at 3+ messages) are listed in the paper's appendix.
Does CreaMem win just by feeding the model more context?
No. The paper reports CreaMem uses comparable or fewer tokens than structured baselines (about 7,250 on LongMemEval-S, roughly an order of magnitude below full-history), and its ablations attribute the gain to dual episodic/trait encoding rather than context size - the RRF fusion and balanced-sampling components show statistically null effects.
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 2 and 3 chart values are extracted as bare numeric series with interleaved layout; series-to-configuration mapping was cross-checked against in-text numbers (75.1 full, 65.9 episodic-only, 2.3-point scene gap), which all match.
- Figure 1 (architecture diagram) has no extractable content; the architecture was assessed from Sec. 3 descriptions and the Appendix B prompt templates.
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.