MemForest: Efficient Agent Memory Management via EventTree Partitioning and Progressive Merging
Junxi Wang, Te Sun, Jiayi Zhu, Chen Zhang, Siyuan Li, Xuyang Liu, Zichen Wen, Xiaobing Tu, et al.
cs.AI
Quick verdict
Claims 50% agent-memory compression at 97-99.7% retention with speedups; broad results, but no significance tests and AGPR gains confound attribution.
Summary
MemForest is a post-processing framework that compresses stored agent memory to roughly half its size while preserving answer quality. KMeans clustering over embeddings yields initial event units, re-partitioned by fusing global similarity-to-cluster-centers with a binary local temporal-continuity score; each unit becomes a maximum spanning tree (EventTree) whose edge weights are modulated by node degrees, and the highest-weight pair is iteratively merged by GPT-4o-mini (summary re-embedded, timestamp set to the max) until the target ratio. An optional Anchor-Guided Propagation Retrieval (AGPR) reranks L·k candidates by temporal proximity to λ·k anchor nodes. Under Mem0 across LoCoMo, LongMemEval, and PersonaMem, MemForest alone retains 97.1% of uncompressed accuracy at 50% compression (ToMe 93.5%, random pruning 82.3%); under M3-Agent on two M3-Bench benchmarks it retains 99.7%. Retrieval speedups are reported for the full system with AGPR: 1.89x (Mem0) and 2.24x (M3-Agent) averages. AGPR alone lifts uncompressed M3-Agent to 117.6% of baseline and adds 21.0 points at 50% compression, attributed to the weaker base retriever. Key limitations: three-run averages without variance or significance tests under GPT-4o-mini judging; AGPR is never combined with baseline compressors, so 120%+ headline figures conflate retrieval and compression gains; the Section 4 bound assumes embedding-interpolation merging rather than the actual LLM re-embedding; dedicated memory-consolidation baselines (LightMem, SimpleMem) are cited but untested.
Review complete
Assessment at a glance
6
risk flags identified
● high 2
● medium 2
● low 2
4 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 event units requires combining global semantic similarity with local temporal continuity, since either signal alone misgroups nodes (Fig. 1b, Table 3).
- At 50% memory compression, MemForest without AGPR retains 97.1% of uncompressed accuracy under Mem0 across LoCoMo, LongMemEval, and PersonaMem, and 99.7% under M3-Agent across M3-Bench-robot and M3-Bench-web.
- The full system with AGPR achieves average retrieval speedups of 1.89x (Mem0) and 2.24x (M3-Agent) at 50% compression (Tables 13-14).
- Merging high-similarity node pairs first provably bounds query-similarity loss by (1−λ)√(2−2ρij) (Eq. 15), and maximum-spanning-tree-guided merging beats random and minimum-spanning-tree merging by 3-7 points (Tables 4, 6, 8).
- AGPR consistently improves retrieval, with much larger gains on M3-Agent (+17.6 points uncompressed, +21.0 at 50% compression) than Mem0 (+0.2/+1.2), attributed to M3-Agent's weaker base retriever and higher multimodal redundancy.
Strengths
Well-motivated partitioning criterion with empirical support
Fig. 1b quantifies the two motivating signals on LoCoMo (m1-m9 similarity 0.73 across distant windows; m1-m2 only 0.12 within one event), and Table 3 shows removing either score costs 2.4-4.6 retention points at 50% compression (96.0% vs 91.5%/93.6%).
Unusually complete ablations and cost transparency
Tables 3-4 (partitioning/merging variants at three ratios, D.1), Table 9 (degree term, +0.5-1.0 points), Figs. 5-6 (all seven hyperparameters), D.5 ($0.10 per LoCoMo 50% compression, 6,000 nodes), D.6 (merging-model sweep: Qwen2.5-7B 87.9% vs GPT-4o-mini 97.0% vs GPT-5.2 96.5% retention).
Honest headline metrics
Arithmetically consistent efficiency reporting
Weaknesses
No variance or significance testing
AGPR gains are attributionally confounded
Theory-practice gap in Section 4
Baseline set misses the relevant competitors
Reproducibility gaps and possible tuning-on-test
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
MemForest (arXiv 2609.08273, cs.AI, Sep 2026) is a post-processing compression framework for agent memory, motivated by Fig. 1's measurement that memory-node growth degrades Mem0 retrieval latency. It contributes three pieces: EventTree Semantic-Temporal Partitioning (§3.2: KMeans over embeddings with α=0.05, then reassignment by fusing global similarity-to-centers with a binary local temporal score, β=0.8, window w=5), EventTree Progressive Merging (§3.3: Kruskal maximum spanning tree per unit, degree-adjusted weights η=0.99, iterative GPT-4o-mini merges with timestamp=max until the target ratio), and Anchor-Guided Propagation Retrieval (§3.4: rerank L·k=4k candidates by fused similarity and temporal proximity to λ·k anchors, γ=0.9). Central claims: at 50% compression MemForest alone retains 97.1% (Mem0) and 99.7% (M3-Agent) of uncompressed accuracy with 1.89x/2.24x retrieval speedups; and Eq. 15 bounds query-similarity loss when merging similar nodes first.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 4 strengths & 5 weaknesses, analyzed in detail
- 6 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 MemForest and what problem does it solve?
MemForest is a post-processing framework that compresses the long-term memory an LLM agent has already stored. It groups memory entries into event-based tree structures (EventTrees) and merges similar entries using an LLM, cutting memory size roughly in half while keeping the agent's answer accuracy nearly unchanged and making retrieval faster.
How much accuracy does MemForest lose at 50% compression?
According to the paper, without the optional AGPR retrieval add-on it keeps 97.1% of the uncompressed accuracy under the Mem0 framework (averaged over LoCoMo, LongMemEval, and PersonaMem) and 99.7% under the multimodal M3-Agent framework (M3-Bench-robot and M3-Bench-web). Losses grow at 70% compression, where Mem0 retention drops to 93.3%.
What is Anchor-Guided Propagation Retrieval (AGPR) in MemForest?
AGPR is an optional retrieval step: it fetches a larger candidate pool (4x the final k), designates the most similar 20% as anchor nodes, and boosts candidates that lie temporally close to those anchors. Its largest effect is on M3-Agent, where even without any compression it raises scores to 117.6% of the original baseline; on Mem0 the effect is small (+0.2 points).
Is MemForest code and data available, and what does it cost to run?
The paper states code is available at github.com/Celina-love-sweet/MemForest. Experiments use GPT-4o-mini for merging and judging (GPT-4o for M3-Agent evaluation) and OpenAI text-embedding-3-small/large embeddings; the authors report the merging cost at about $0.10 for compressing 50% of the LoCoMo benchmark (about 6,000 memory nodes).
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.
- Mathematical notation inside Figure 2 and some display equations (e.g., Eq. 9's exponent) is garbled by PDF extraction with placeholder glyphs; symbol-level judgments (such as whether absolute-value bars are present) were deliberately avoided rather than assumed.
- The provided metadata lists 8 authors, while the extracted full text lists 13 (adding Jinkui Ren, Xiantao Zhang, and Ziqi Yuan with Alibaba/Tsinghua affiliations).
- The tables' bold 'top-two' highlighting described in captions is lost in extraction; rankings cited in this review were computed from the printed numbers, not the formatting.
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.