JustMem: Just-Enough Memory Access for Long-Term Conversations
Guanhua Chen, Yanting Wang, Wenjing Zhi, Lei Sha
cs.CL
Quick verdict
Proposes query-adaptive memory access (LOOKUP/COMPOSE/REPLAY); strong efficiency-accuracy results, but LoCoMo edge over LightMem is not significant (p=0.53).
Summary
JustMem is a long-term conversational memory system that compresses conversation history into compact atomic memory cards with provenance links and adapts memory access to each query along two dimensions: discovery breadth and reading fidelity. A single planner call predicts entities, temporal constraints, operations, and two binary access demands that deterministically route each query to LOOKUP (narrow discovery, compact reading), COMPOSE (multi-rewrite candidate union under a fixed Top-10 budget), or REPLAY (provenance-linked source recovery). Using GPT-4.1-mini for extraction, planning, answering, and judging, JustMem attains 79.61% LLM-judge accuracy on LoCoMo and 83.40% with 96.00% Recall@10 on LongMemEval-S — the best mean accuracy and recall among six compared systems — while cutting construction and inference tokens by roughly 50–67% versus LightMem. Controlled ablations support each design choice: global atomic retrieval beats session-level organization, guided expansion helps distributed-evidence queries most, and query-gated recovery matches fixed recovery on fidelity-sensitive questions at a fraction of the cost. On the LoCoMo-Plus cognitive split, JustMem shows the smallest performance drop (17.02 points). Main limitations: the LoCoMo advantage over LightMem is not statistically significant (p=0.532); reported variance excludes extraction and routing stochasticity; only a single proprietary backbone is tested with an LLM judge; and no code release is stated.
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
- Long-term conversational memory access should be allocated per query along two dimensions, discovery breadth and reading fidelity, rather than fixed globally.
- JustMem achieves the highest mean accuracy and Recall@10 among compared memory systems on LoCoMo and LongMemEval-S while using the smallest generative budgets (~50–67% fewer construction/inference tokens than LightMem).
- Global atomic competition (H1), query-guided discovery breadth (H2), and query-gated source recovery (H3) each causally improve performance under controlled comparisons.
- Adaptive routing approaches benchmark-oracle routing quality (83.40 vs. 84.20 on LongMemEval-S) and transfers best to the LoCoMo-Plus cognitive split (gap of 17.02 points vs. 33.45–49.71 for baselines).
Strengths
Clean conceptual formulation of access allocation
Separating discovery breadth from reading fidelity (§3, Figure 2) is a crisp abstraction that contrasts with representation-focused prior work (A-Mem, MemoryOS, SimpleMem); Eq. 1 maps two predicted binary demands to three configurations deterministically with no extra model call.
Matched-configuration comparison protocol
§4.3 holds GPT-4.1-mini for construction/answering/judging, text-embedding-3-small, dataset-specific prompts constant across all six systems, isolating memory representation and access strategy rather than model choice.
Efficiency is the headline, and it holds
Well-controlled ablations with paired statistics
Diagnostic REPLAY decomposition
Weaknesses
LoCoMo accuracy superiority is not statistically supported
Variance protocol understates end-to-end uncertainty
Single proprietary backbone and unvalidated LLM judge
Transfer comparison (Table 6) omits the strongest baselines
Reproducibility gaps
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
JustMem (Beihang University; arXiv 2609.19877) is a method paper on long-term conversational memory for LLM assistants. Its central claim is that memory *access* — not just memory representation — should be adapted per query along two dimensions: discovery breadth (how widely evidence is searched) and reading fidelity (compact card vs. original conversation). Three query-conditioned configurations instantiate this: LOOKUP (narrow discovery, compact reading), COMPOSE (multi-rewrite candidate union under a fixed Top-10 budget), and REPLAY (provenance-linked source recovery). The authors claim the highest mean accuracy and Recall@10 among compared systems on LoCoMo and LongMemEval-S at the smallest generative budgets, validated by three controlled hypotheses (H1–H3, §5.2–5.5) and a transfer study on LoCoMo-Plus (§5.6).
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 JustMem and what problem does it solve?
JustMem is a memory system for LLM assistants that must answer questions using many past conversation sessions. Instead of applying one fixed retrieval strategy to every question, it predicts whether the needed evidence is local, spread across sessions, or detail-sensitive, then reads compact memory summaries, searches more broadly, or pulls back the original conversation text accordingly — keeping the context small unless the query demands more.
How is JustMem different from standard RAG?
Standard RAG retrieves chunks with a single fixed strategy and fixed context size. JustMem stores atomic memory cards with provenance links to the original conversation and routes each query to one of three configurations: LOOKUP (local evidence), COMPOSE (broader search using planner-generated query rewrites under the same top-10 evidence budget), or REPLAY (recovering original conversation content when compact summaries may have lost answer-critical detail).
What results does JustMem report on LoCoMo and LongMemEval?
On LoCoMo it reports 79.61% LLM-judged accuracy and 88.05% Recall@10; on LongMemEval-S, 83.40% accuracy and 96.00% Recall@10 — the best means among the compared systems — while using the fewest generative tokens (about 50–67% fewer construction and inference tokens than LightMem). On the LoCoMo-Plus cognitive split, its accuracy drop from LoCoMo is the smallest (17.02 points). Note the LoCoMo accuracy edge over LightMem is not statistically significant (p=0.532).
Is JustMem's code publicly available?
The reviewed text does not mention a public code release. Reimplementation is nonetheless feasible: the datasets (LoCoMo, LongMemEval-S, LoCoMo-Plus), the backbone (GPT-4.1-mini), the embedder (text-embedding-3-small), the reranker (Qwen3-Rerank), and condensed prompts for extraction, planning, and answering (Appendix A) are all described in the paper.
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 (1–3) are available only through extracted captions and surrounding text; visual details could not be verified.
- Appendix A prompts are explicitly condensed by the authors; full prompts, formatting constraints, and validation instructions are not in the text.
- Hyperparameters of the retrieval scoring (lexical/temporal/card-aware signal weights, rerank thresholds) are absent from the extracted text.
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.