acceptodds
Under review as a conference paper at ICLR 2027

MemFolder: Structure-Guided Memory Folding for Efficient Agent Memory Management

Abstract

Long-term agents accumulate large external memory stores as interactions unfold, increasing the costs of storage, retrieval, and response generation. Efficiency is constrained by two coupled factors: a growing record count increases retrieval overhead, while longer retrieved contexts increase prefill computation and decoding-time attention costs. Existing methods prune or consolidate records, but pruning can discard useful evidence, whereas consolidation may produce fewer yet longer retrieval units. We introduce MemFolder, a structure-guided memory folding framework that sequentially reduces record count and context length. The framework comprises two core components that share memory groups, using cross-record structure to guide both evidence consolidation and compact representation. RecordFolder grounds entity mentions, progressively selects representative yet complementary anchors, and combines semantic and entity affinity with temporal chaining to consolidate compatible records into fewer retrieval units. ContextFolder encodes the original records within these groups into compact text and uses verification feedback to revise encodings when key facts cannot be recovered. Experiments on the three text-based memory benchmarks LoCoMo, PersonaMem, and LongMemEval and the two multimodal benchmarks M3-Bench-web and M3-Bench-robot demonstrate the strong performance of our methods. Notably, on Mem0, RecordFolder retains 99.0% of the original average performance across the three text-based benchmarks with 50% fewer records, exceeding the strongest baseline by 1.9 percentage points in performance retention. Combined with ContextFolder, existing record compression methods further shorten their output records while retaining most of their performance.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Agent Memory, Large Language Models, Memory Management

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_BjJmOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes MemFolder, which combines record-level consolidation (RecordFolder) and context-level compression (ContextFolder) to improve agent memory efficiency. Experiments on text-based and multimodal benchmarks show promising results, including 99.0% performance retention with 50% fewer records on Mem0. However, the novelty over existing memory compression methods, the reliability of factual verification, and the end-to-end efficiency gains remain insufficiently established.

Critical strengths
  1. The paper addresses two distinct bottlenecks in agent memory management: the number of stored records and the length of retrieved contexts. Treating them separately is a sensible design choice, particularly for agents that need to maintain long interaction histories under limited memory and context budgets.
  2. RecordFolder and ContextFolder target different sources of memory overhead, making their combination more comprehensive than record pruning or context compression alone. The evaluation across text-based and multimodal settings also provides initial evidence that substantial memory reduction is possible without a proportional drop in downstream performance.
Critical weaknesses
  1. The novelty over existing memory compression methods remains unclear. MemForest, MemRefine, and StreamMeCo have explored related directions, including memory consolidation and compression [1–3]. Although MemFolder combines record-level and context-level compression, the methodological advance beyond these approaches is not sufficiently established. The relatively small contribution of some components in the ablations further weakens the novelty claim.
  2. The evidence for factual reliability and practical efficiency is incomplete. ContextFolder relies on an LLM-based verification loop, yet the paper acknowledges inconsistent verifier judgments, and facts omitted during extraction may never be checked. In addition, the reported speedups mainly concern retrieval, while preprocessing also incurs LLM and verification costs. Consequently, neither reliable fact preservation nor the overall deployment benefits are fully established. [1] Wang et al. MemForest: Efficient Agent Memory Management via EventTree Partitioning and Progressive Merging. arXiv:2609.08273, 2026. [2] Kim et al. MemRefine: LLM-Guided Compression for Long-Term Agent Memory. arXiv:2606.13177, 2026. [3] Wang et al. StreamMeCo: Long-Term Agent Memory Compression for Efficient Streaming Video Understanding. Findings of ACL 2026.
Questions for authors
  1. A controlled comparison with MemForest and MemRefine under comparable memory budgets would help clarify the methodological differences and the additional value of MemFolder.
  2. Reporting whether preprocessing costs, including LLM calls and factual verification, are included in the latency evaluation would provide a clearer picture of the overall efficiency gains.
  3. An independent factual evaluation, including cases where important facts are omitted or incorrectly preserved, would provide stronger evidence of the reliability of ContextFolder.
7 comments
AuthorsOpenReview ↗

Response to Reviewer BjJm (1/5: Weakness 1 - Methodological Advance)

Thank you very much for your constructive review and for recognizing that decoupling stored record count from retrieved context length is a sensible design choice, that RecordFolder and ContextFolder target complementary overhead sources more comprehensively than pruning or context compression alone, and that our text and multimodal evaluations demonstrate substantial memory reduction without a proportional performance drop. Below, we respond to each of your weaknesses and questions in detail:

Weakness 1: Methodological Advance Over Existing Compression Methods (Part 1: Mechanisms & Macro-Average Results)

Unlike existing methods that address a single bottleneck—either losing evidence via pruning or inflating unit length by 2.12x–3.87x via merging—MemFolder introduces a reusable shared grouping structure that decouples and sequentially optimizes record count N and per-unit length l.

1. Architectural Comparison with StreamMeCo, MemRefine, and MemForest

DimensionStreamMeCo (ACL'26 Findings)MemRefine (arXiv 2026)MemForest (arXiv 2026)MemFolder (Ours)
Optimization & BudgetSingle-axis (node count N only); prunes graph nodes without shortening textCoupled byte budget; iterates over JSON-serialized length without separate K and l controlsSingle-axis (record count N only); merging inflates unit length by 2.12–3.01xDual-axis decoupled: RecordFolder sets K = max{1, round(kappa N)}; ContextFolder cuts unit tokens by 31%–40%
Grouping & StructureRepresentative sampling + importance/similarity pruningLocal cosine-similarity candidate pairs judged via {delete, merge, preserve}Maximum spanning tree (EventTree) with progressive pairwise edge mergingCross-session entity registry + DPC progressive anchors + rare-entity IDF affinity + Hungarian chains
Merging & RewritingKeeps unedited verbatim nodes (deleted evidence is lost)Local iterative pairwise merge/delete without global coverage constraintsMulti-step pairwise LLM summarization along tree edges; no compact syntaxSingle-pass time-ordered group merge (0 calls for singletons) + syntax-guided encoding of source lines
Factual VerificationNoneJudge decides merge/delete only; no blind reader recovery checkNoneFact-importance planning + blind reader QA recovery check + targeted revision & fallback
ComposabilityDedicated graph pruning pipelineClosed single-loop iterationCoupled with AGPR in original paper (unified retriever used in our tests)ContextFolder plugs into MemForest/ToMe to further compress their outputs

2. Macro-Average Accuracy Under Fixed Record-Count Budgets (Tables 1–3 Summary) Within each table and figure, all methods are evaluated on identical memory snapshots, retrievers (fixed top-k), and QA protocols. RecordFolder achieves the highest macro-average accuracy across every evaluated framework and reduction ratio:

Framework & ScopeReduction (Retention)Best Pruning BaselineToMeMemForestRecordFolder (Ours)Gain vs. MemForest
Mem0 (3 text benchmarks)Uncompressed (0%)61.1 (100.0%)61.1 (100.0%)61.1 (100.0%)61.1 (100.0%)—
30% (Keep 70%)58.2 (95.3%, KMeans)58.3 (95.4%)59.7 (97.7%)61.5 (100.7%)+1.8 / +3.0 pp
50% (Keep 50%)53.0 (86.7%, KMeans)57.1 (93.5%)59.3 (97.1%)60.5 (99.0%)+1.2 / +1.9 pp
70% (Keep 30%)45.3 (74.1%, DART)56.0 (91.7%)57.0 (93.3%)59.0 (96.6%)+2.0 / +3.3 pp
MemOS (3 text benchmarks)Uncompressed (0%)71.1 (100.0%)71.1 (100.0%)71.1 (100.0%)71.1 (100.0%)—
50% (Keep 50%)56.7 (79.7%, KMeans)63.6 (89.4%)63.6 (89.5%)64.7 (91.0%)+1.1 / +1.5 pp
70% (Keep 30%)47.2 (66.4%, KMeans)58.7 (82.6%)58.7 (82.6%)60.3 (84.8%)+1.6 / +2.2 pp
M3-Agent (Robot & Web)Uncompressed (0%)39.1 (100.0%)39.1 (100.0%)39.1 (100.0%)39.1 (100.0%)—
50% (Keep 50%)37.7 (96.2%, StreamMeCo)38.5 (98.5%)39.0 (99.7%)39.6 (101.3%)+0.6 / +1.6 pp
70% (Keep 30%)35.1 (89.8%, StreamMeCo)38.2 (97.7%)38.5 (98.5%)40.1 (102.6%)+1.6 / +4.1 pp

In Part 2/5, we address component ablations and Question 1's dual-axis budget comparisons.

AuthorsOpenReview ↗

Response to Reviewer BjJm (2/5: Weakness 1 - Ablations & Question 1)

Weakness 1 (Continued): Component Ablation Analysis

The ablation results in Tables 4–7 and Figure 5B reflect proximity to the uncompressed baseline and targeted gains on cross-session, preference, and knowledge-update subtasks:

  • 1. Structured grouping over standard clustering (Tables 4 & 5): At 50% record reduction, DPC achieves the highest average score among tested anchor strategies (57.2, beating KMeans 55.4 by +1.8 and runner-up MinMax 56.8 by +0.4, retaining 99.1% of the uncompressed 57.7 baseline), while removing cosine similarity drops accuracy by 3.4 points. Operating near the uncompressed baseline, gain gating (+1.9), entity affinity (+1.1), and temporal chaining (+0.6) act as fine-grained structural refinements.
  • 2. Substantial gains on targeted LongMemEval subtasks (Tables 1 & 2): Overall averages are diluted by single-hop questions requiring one record. On subtasks directly testing cross-session entity resolution and temporal state updates, RecordFolder achieves pronounced gains over MemForest consistent with our design motivation: in the full record-compression evaluation on Mem0 (Table 1, 50% reduction), it leads overall by +3.2 points (55.0 vs. 51.8; +5.3 on Multi-Session MS, +10.0 on Preference SP, +3.0 on Temporal TR) and by +30.0 points on SP at 70% reduction; on MemOS (Table 2, 50% reduction), it leads by +16.6 points on SP (83.3 vs. 66.7), +10.2 points on Knowledge Update KU (75.6 vs. 65.4), and +4.5 points on MS (45.1 vs. 40.6).
  • 3. Clear value of planning and blind verification (Table 7 & Figure 5B): Removing Planner increases unit length by +32.0% (12.31 -> 16.25 tokens on LoCoMo) and reduces accuracy by 0.7–0.8 points; removing Verifier drops accuracy by 1.9 points (LoCoMo) and 1.4 points (LongMemEval) at slightly lower achieved tokens (11.94 vs. 12.31). Increasing revision iterations from 1 to 3 (Figure 5B) boosts accuracy by +2.14 points on LoCoMo (54.42% -> 56.56%) and +3.00 points on LongMemEval (48.80% -> 51.80%) at nearly identical token counts.

Question 1: Controlled Comparison Under Comparable Memory Budgets

In the context-compression evaluation at 50% record reduction (Figure 3, tabulated below), full MemFolder operates in a compact ~10.8–13.4 tokens/record regime, outperforming MemForest+ContextFolder by +2.8 points on LoCoMo and LLMLingua-2 by up to +12.2 points, while ContextFolder cuts MemForest and ToMe outputs by 23.2%–42.6%:

Record Method (Keep 50%)Context MethodLoCoMo Tok/RecLoCoMo ACCLongMemEval Tok/RecLongMemEval ACCM3-Robot Tok/RecM3-Robot ACC
RecordFolder (Ours)Vanilla20.6059.321.6055.230.231.0
+ BabelTele27.30 (+32.5%)49.317.20 (-20.4%)45.830.2 (0.0%)32.1
+ LLMLingua-210.80 (-47.6%)44.412.00 (-44.4%)47.831.7 (+5.0%)29.5
+ ContextFolder (Ours)12.31 (-40.2%)56.613.40 (-38.0%)51.820.9 (-30.8%)29.5
MemForestVanilla19.1058.620.8054.630.431.6
+ BabelTele27.00 (+41.4%)48.416.90 (-18.8%)45.029.3 (-3.6%)31.5
+ LLMLingua-210.20 (-46.6%)42.511.40 (-45.2%)48.231.8 (+4.6%)29.9
+ ContextFolder (Ours)11.50 (-39.8%)53.813.30 (-36.1%)52.220.7 (-31.9%)27.4
ToMeVanilla22.3055.719.8052.631.730.9
+ BabelTele30.00 (+34.5%)46.517.10 (-13.6%)47.830.6 (-3.5%)30.6
+ LLMLingua-211.90 (-46.6%)36.911.00 (-44.4%)47.431.8 (+0.3%)31.4
+ ContextFolder (Ours)12.80 (-42.6%)51.815.20 (-23.2%)51.221.5 (-32.2%)29.0
  • Controlled comparison with MemRefine's mechanism: MemRefine [2] operates on multi-field structured notes (keywords, tags, context descriptions) under a coupled JSON-serialized byte budget using local pairwise {delete, merge, preserve} actions without a dedicated intra-group syntax encoder. Under identical anchors and storage budgets on our conversational stores, Table 6 (Merging 57.2 vs. Pruning 51.1, a +6.1-point advantage) and Figure 3 above (where adding verified compact encoding cuts unit tokens by 31%–40% while outperforming token-deletion baselines by up to +12.2 points) directly isolate and confirm the advantage of MemFolder's decoupled consolidation and verified encoding over deletion-reliant compression.
AuthorsOpenReview ↗

Response to Reviewer BjJm (3/5: Weakness 2 - Factual Reliability)

Weakness 2 (Part 1): ContextFolder Factual Reliability, Trace Audit, and Held-Out Utility

While we explicitly frame ContextFolder's verification loop as an asymmetric failure-detection heuristic rather than a formal lossless guarantee (Section 3.4), dual-input access to source records, reference-aligned judging, and length-safe fallbacks measurably protect factual fidelity in both source-level production traces and held-out QA evaluations.

1. Defense Mechanisms and Residual Boundaries Across Four Factual Failure Modes (Section 3.4 & Figures 9–14)

Failure ModeReviewer's ConcernBuilt-in Safeguard in ContextFolderObserved Behavior & Residual Boundary
1. Extraction OmissionFacts omitted by Planner are never checked by VerifierDual-input Compressor (Fig. 10): Receives both {facts_list} and complete raw group records {lines}. Adaptive density extraction (Fig. 9, ~1 fact/10 tokens, up to 12) prioritizes high-tier facts.Prevents Planner omissions from cutting off source access; verification coverage remains bounded by extracted checks, and low-priority modifiers in >12-fact groups are unchecked.
2. Compression OmissionCompressor drops dates, negations, or key facts for token budgetTiered retention (Fig. 10): Mandates non-inferable high-tier facts (and med-tier when tau >= 0.65) with exact names/dates/numbers. Blind reader check (Fig. 11): Forces "UNKNOWN" when missing, triggering MUST-FIX revision (Fig. 13).Prioritizes shortest passing candidate; if none passes in 3 rounds, selects by weighted recovery (3:2:1) and falls back to a shorter candidate (below S_G) or raw concatenation if length >= S_G.
3. Substitution ErrorBlind reader answers a missing fact using another fact in the encodingReference-aligned Fact Judge (Fig. 12): Compares blind answers against Expected Answers, requiring exact numbers, dates, and entity names; mismatches fail and trigger revision.Catches "same-category fact substitution" that a simple "UNKNOWN" check would miss (Case C below).
4. Verifier InconsistencySame encoding receives inconsistent judgments across rounds (App. B.4)Asymmetric isolation (App. B.4): Blind reader isolation prevents direct answer copying from source text; conservative "UNKNOWN" outputs prompt clearer phrasing in subsequent rounds.While Fact Judge's lenient semantic matching does not provide a zero-false-pass formal guarantee, multi-round revision and fallback protection substantially reduce missed facts in practice.

2. Source-Level Trace Audit (Appendix B.1–B.3) Across the three unmodified LoCoMo production traces in Appendix B (22 source records totaling 202 tokens -> 88 tokens, -56.4%), 13/13 Planner-extracted checks passed within 1–3 rounds:

  • Intercepting Fact Substitution (App. B.3, Case C): In a 12-record group (92 tokens), Round 1 (44 tokens) omitted Fact 3 ("Values counseling and support groups"), and the blind reader substituted Fact 6 ("sharing stories and giving others hope"). Fact Judge caught this mismatch against the expected answer and flagged Fact 3 for MUST-FIX; Round 2 (50 tokens, 0.54 ratio) restored it and passed all 6 checks.
  • Resolving Coreference Ambiguity (App. B.1, Case A): In a 4-record group (48 tokens), Rounds 1 (19 tokens) and 2 (26 tokens) failed Question 4 ("UNKNOWN"); Round 3 (21 tokens, 0.44 ratio) passed all 4 checks (weighted recovery 1.0) and was selected over the shorter incomplete Round 1.
  • Single-Round High-Redundancy Folding (App. B.2, Case B): In a 6-record group (62 tokens), Round 1 passed all 3 checks in 17 tokens (0.27 ratio, -72.6% tokens).

3. Held-Out Downstream QA Ablations (Zero Test-QA Input)

MethodPlannerVerifierLoCoMo Tok/RecLoCoMo ACC (%)LongMemEval Tok/RecLongMemEval ACC (%)
RecordFolder + LLMLingua-2NoNo10.8044.4 (-12.2)12.0047.8 (-4.0)
RecordFolder + BabelTeleNoNo27.3049.3 (-7.3)17.2045.8 (-6.0)
ContextFolder (w/o planner, Table 7)NoYes (3 rounds)16.2555.9 (-0.7)14.4451.0 (-0.8)
ContextFolder (w/o verifier, Table 7)YesNo (0 rounds)11.9454.7 (-1.9)13.1950.4 (-1.4)
ContextFolder (Iteration = 1, Fig. 5B)Yes1 pass (no rev.)11.9454.42 (-2.14)13.0848.80 (-3.00)
Full ContextFolder (Ours, 3 rounds)YesYes (up to 3)12.3156.5613.4051.80

At comparable achieved token lengths (LoCoMo: 11.94–12.31; LongMemEval: 13.08–13.40 tokens/record), 3-round verification improves downstream accuracy by +1.9 and +1.4 points over w/o verifier, and by +2.14 and +3.00 points over single-pass Iteration = 1.

Reviewer_33d6OpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

this paper proposes MEMFOLDER to compress records into managed memory. The framework includes RecordFolder to cluster and merge, and ContextFolder to compress records. Empirical results show MEMFOLDER outperform competitive baselines

Critical strengths

The paper addresses a practically important problem in long-term agent memory management and provides extensive empirical evaluations across multiple benchmarks and memory frameworks. The proposed approach achieves favorable accuracy–efficiency trade-offs, consistently outperforming strong memory-compression baselines under different compression ratios.

Critical weaknesses

The methodological novelty of MemFolder appears limited, as several closely related works have already explored its central ideas:

  • SimpleMem (2026) combines semantic compression and recursive consolidation of related memories, overlapping with MemFolder's overall two-stage approach.
  • LightMem (ICLR 2026) uses hierarchical clustering to organize memories, raising questions about the necessity of MemFolder's more elaborate grouping strategy.
  • REMem (ICLR 2026) connects semantically related episodic memories and maintains compact gists alongside structured facts, overlapping with MemFolder's entity-aware memory organization and compact representations.

While the paper explicitly distinguishes RecordFolder from MemForest through progressive anchor selection and entity-aware temporal chaining, it remains unclear whether these algorithmic differences yield meaningful improvements.

Questions for authors

While the paper reports substantial retrieval speedups, MemFolder introduces additional LLM calls for entity grounding, consolidation, planning, and iterative verification. What are the total preprocessing costs, and how many downstream queries are needed to amortize them? Does the full pipeline improve end-to-end latency and cost compared with MemForest and simpler baselines?

6 comments
AuthorsOpenReview ↗

Response to Reviewer 33d6 (1/4: Weakness 1 - Architectural Comparison)

Thank you for your thoughtful review and for recognizing that long-term agent memory management is a practically important problem, that our work provides extensive empirical evaluations across multiple benchmarks and memory frameworks, and that MemFolder achieves favorable accuracy–efficiency trade-offs and consistently outperforms strong baselines across compression ratios. Below, we address each of your comments in detail:

Weakness 1: Distinction from SimpleMem, LightMem, and REMem (Part 1: Architectural Comparison)

Unlike SimpleMem, LightMem, and REMem—which are end-to-end specialized memory architectures—MemFolder is a post-construction dual-axis folding layer that operates on already built memory stores to simultaneously reduce record count and per-unit token length.

We discussed SimpleMem in Section 2 of the original submission and have added detailed discussions of LightMem and REMem in Section 2 of the revised manuscript. While these methods explore compression, clustering, or structured indexing during memory construction, standard record merging alone inflates per-unit length by 2.12x–3.87x on LoCoMo (Figure 1). MemFolder resolves this bottleneck via a reusable shared grouping structure that couples intra-group record consolidation (RecordFolder) with budget-guided compact encoding and blind recovery verification (ContextFolder):

DimensionSimpleMem (2026)LightMem (ICLR 2026)REMem (ICLR 2026)MemFolder (Ours)
Core RoleEnd-to-end lifelong memory (sliding-window gating + online recursive synthesis)Three-stage cognitive memory (sensory filter + topic STM + sleep-time LTM update)Hybrid episodic graph (time-aware Gists/Facts graph + multi-step agentic retrieval)Post-construction dual-axis folding layer compatible with existing stores
Grouping MechanismLocal semantic similarity threshold; no cross-session entity registry or non-branching chainsTopic segmentation and embedding clustering during sleep-time consolidationExtracts Gists and temporal Facts into a hybrid graph; does not merge or reduce nodesCross-session entity registry + DPC progressive anchors + Hungarian non-branching chains
Dual-Axis Reduction (N & l)No (recursive merging inflates node length by 1.57x at 50% reduction)No (sleep-time merging inflates node length by 2.17x at 50% reduction)No (stores Verbatim + Gists + Facts simultaneously, expanding node count)Yes (halves record count via RecordFolder and cuts unit tokens by 31%–75% via ContextFolder)
Factual VerificationNoneSingle-pass offline LLM rewrite without blind recovery checkNone (relies on multi-step online agent tool calls)Fact-importance planning + controlled syntax encoding + blind reader verification & fallback

In the next comment (Part 2/4), we present controlled experiments applying MemFolder directly on top of pre-built SimpleMem, LightMem, and REMem stores.

AuthorsOpenReview ↗

Response to Reviewer 33d6 (2/4: Weakness 1 - Plug-and-Play Experiments)

Weakness 1: Distinction from SimpleMem, LightMem, and REMem (Part 2: Post-Construction Dual-Axis Synergy)

When applied on top of pre-built SimpleMem, LightMem, and REMem stores at 50% record retention, full MemFolder simultaneously outperforms native consolidation in both accuracy (+5.32 to +11.10 points) and per-node length (-23.3% to -54.5%).

To verify orthogonality, we evaluated ① each framework's native consolidation/clustering (Native@50%), ② RecordFolder at 50% retention (RF@50%), and ③ full MemFolder combining RecordFolder and ContextFolder (RF + CF, MemFolder@50%) under identical snapshots, retrievers, and QA evaluators:

BenchmarkBase FrameworkCompression SettingQA ACC (%)Tokens/NodeLength Change (vs. Full / Native)
LoCoMoSimpleMemFull Context (Uncompressed)74.4829.7—
+ Native Consolidation@50%60.1346.5+56.6% (inflated vs. Full)
+ RecordFolder@50% (Ours)74.2973.8+148.5% (retains full details)
+ MemFolder@50% (Ours, RF+CF)65.4524.4-17.8% (-47.5% vs. Native)
LightMemFull Context (Uncompressed)76.3023.4—
+ Native Consolidation@50%59.6150.8+117.1% (inflated vs. Full)
+ RecordFolder@50% (Ours)75.1958.9+151.7% (retains full details)
+ MemFolder@50% (Ours, RF+CF)70.7123.1-1.3% (-54.5% vs. Native)
REMemFull Context (Uncompressed)81.3634.1—
+ Native Consolidation@50%N/A (no post-store merging)——
+ RecordFolder@50% (Ours)83.6465.8+93.0%
+ MemFolder@50% (Ours, RF+CF)77.5316.6-51.3% (halved vs. Full)
LongMemEvalLightMemFull Context (Uncompressed)75.6019.8—
+ Native Consolidation@50%65.0030.5+54.0% (inflated vs. Full)
+ RecordFolder@50% (Ours)72.0033.4+68.7%
+ MemFolder@50% (Ours, RF+CF)69.5023.4-23.3% vs. Native (+18.2% vs. Full)
SimpleMemFull / Native@50% / RF@50% (Ours)43.80 / 41.40 / 43.9031.0 / 43.0 / 53.9Surpasses Full and Native ACC
REMemFull / RF@50% (Ours)52.90 / 56.2091.3 / 43.2-52.7% (+3.30 ACC gain)

Three key conclusions follow:

  • Simultaneous accuracy and length gains over native consolidation: On LoCoMo at 50% record retention, full MemFolder@50% cuts per-node length by 47.5% on SimpleMem (46.5 -> 24.4) and 54.5% on LightMem (50.8 -> 23.1; and -23.3% vs. native on LongMemEval) while outperforming native consolidation by +5.32 and +11.10 points, respectively. On REMem, RecordFolder@50% boosts accuracy by +2.28 (LoCoMo) and +3.30 points (LongMemEval, cutting tokens from 91.3 to 43.2).
  • Structure-guided grouping preserves cross-session evidence: Native consolidation drops by 14.35 and 16.69 points on LoCoMo, whereas RecordFolder@50% maintains 74.29% (+14.16) and 75.19% (+15.58). While RecordFolder aggregates group details into longer merged units, ContextFolder subsequently eliminates this length inflation.
  • Seamless orthogonal synergy: MemFolder requires zero modifications to underlying extraction or retrieval pipelines.
AuthorsOpenReview ↗

Response to Reviewer 33d6 (3/4: Weakness 2 - Grouping Strategy Necessity)

Weakness 2: Necessity and Empirical Value of Our Grouping Strategy

RecordFolder's progressive density-peaks clustering (DPC) and entity-aware temporal chaining consistently outperform simple clustering (+1.8 to +15.58 points) and MemForest across all frameworks, with the largest gains on cross-session, preference, and knowledge-update subtasks (up to +30.0 points).

Our grouping strategy targets two failure modes of standard embedding clustering (KMeans, LightMem topic clustering) and local spanning-tree merging (MemForest): loss of low-frequency complementary evidence due to embedding crowding, and cross-session coreference or state-evolution breaks:

  • 1. Consistent gains over clustering and MemForest: In the controlled anchor ablation (Table 4, identical downstream merging), replacing Random (54.8) or KMeans (55.4) with RecordFolder's progressive anchors (57.2) improves average accuracy by +1.8 to +2.4 points. In full evaluations (Table 1), RecordFolder beats KMeans by +7.5 and +13.9 points at 50% and 70% reduction on Mem0, and surpasses LightMem's native clustering by +15.58 (LoCoMo) and +7.00 points (LongMemEval) at 50% reduction. Compared with MemForest, RecordFolder achieves the highest macro-average accuracy across all evaluated configurations (Mem0 at 30%/50%/70%; MemOS and M3-Agent at 50%/70%), leading by +2.0 points (96.6% vs. 93.3% retention, +3.3 pp) at 70% reduction on Mem0.
  • 2. Pronounced improvements on cross-session, preference, and knowledge-update subtasks: Overall benchmark averages are diluted by single-hop questions solvable from one record. On LongMemEval subtasks directly aligned with our entity registry and non-branching temporal chain design (Tables 1 & 2)—Multi-Session (MS), Preference (SP), Knowledge Update (KU), and Temporal Reasoning (TR)—RecordFolder achieves substantial gains:
Framework & RetentionLongMemEval SubtaskMemForestRecordFolder (Ours)GainAlignment with Design Motivation
Mem0 (Keep 50%)Overall51.855.0+3.2Driven by multi-session and temporal gains
Multi-Session (MS)48.153.4+5.3Cross-session entity registry resolves coreference
Preference (SP)20.030.0+10.0Progressive anchors protect tail user preferences
Temporal (TR)42.945.9+3.0Non-branching chains preserve event chronology
Mem0 (Keep 30%)Preference (SP)23.353.3+30.0Entity affinity & complementary anchors under high compression
Temporal (TR)39.845.1+5.3Avoids local spanning-tree truncation of long chains
MemOS (Keep 50%)Overall60.064.2+4.2Amplified gains on stronger OS-level memory base
Preference (SP)66.783.3+16.6Complementary anchors preserve preference details
Knowledge Update (KU)65.475.6+10.2Time-ordered chains track sequential state updates
Multi-Session (MS)40.645.1+4.5Rare-entity IDF affinity links cross-session facts
  • 3. Objective ablation framing (Tables 4 & 5): DPC achieves the highest average score among tested anchor strategies (57.2, beating KMeans 55.4 by +1.8 and MinMax 56.8 by +0.4, retaining 99.1% of the uncompressed 57.7 baseline). Near this uncompressed baseline, gain gating (+1.9), entity affinity (+1.1), and temporal chaining (+0.6) provide targeted structural refinements for cross-session and state-update consistency.
Reviewer_CviYOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This work tackles the challenging problem of optimizing memory efficiency and performance in long-horizon agents. Two problems are highlighted specifically: (1) the costs of memory storage as memories are collected and stored over time (2) the computational costs of long contexts. The authors propose a new method, MemFolder, to consolidate memories, reducing record count and context length. Their framework consists of two components. The first, RecordFolder, consolidates compatible records into a smaller number of retrieval units. The second, ContextFolder, encodes records into compact text, using verification feedback to refine encodings. Their experiments on memory benchmarks show promising results for MemFolder.

Critical strengths

-Memory is a popular direction these days - the authors claim that their work reduces two sources of memory inefficiency simultaneously, unlike other works that focus on one. If true, this is a strong differentiator.

-The paper is well structured and easy to follow. The breakdown of RecordFolder and ContextFolder is nice.

-Experimental setup seems reasonable: standard benchmarks and baselines selected. Thorough results including ablations. Latency and efficiency analysis included.

-Results appear promising - MemFolder consistently outperforms baselines on record compression and in context compression, MemFolder has the best accuracy-token tradeoffs compared to baselines.

Critical weaknesses

-Consider simplifying Figure 1. There is a lot going on and it only makes sense after reading the Figure description.

-Same for Figure 2 – consider simplifying, maybe remove the math equations here and try to share the high level ideas only. It’s even more complicated than Figure 1 and hard to read.

Questions for authors

N/A

2 comments
AuthorsOpenReview ↗

Response to Reviewer CviY (W1 & W2)

Thank you very much for your positive review and for recognizing that simultaneously reducing both sources of memory inefficiency is a strong differentiator, that the paper is well structured with a clear breakdown of RecordFolder and ContextFolder, and that MemFolder achieves strong record compression and the best accuracy–token trade-offs across thorough ablations and efficiency analyses. Below, we detail our revisions addressing both of your suggestions:

Weakness 1: Simplifying Figure 1 and Clarifying Metric Scopes

We have redesigned Figure 1 (page 2 of the revised manuscript) so readers can immediately grasp our dual-axis motivation without relying on a lengthy caption:

  • Streamlining the central 2D schematic: In the record-count (N) vs. tokens-per-unit (l) plot, we removed cluttered baseline accuracy numbers, added a clear retrieval-unit legend (= 1 retrieval unit) and failure-zone labels (units too long; N not reduced), and contrasted method behaviors directly: pure pruning drops evidence (Prune: drops evidence), pure merging inflates unit length (Merge: 2.1–2.5x longer), and prompt compression leaves record count unchanged (Prompt compression: N not reduced), whereas MemFolder sequentially folds both axes (fixes both).
  • Clarifying statistical scopes: In the right-hand overview, we retained only the primary quantitative summary (RecordFolder preserves 99.0% average performance across three Mem0 text benchmarks at 50% record reduction; ContextFolder further cuts tokens by 40.2% on LoCoMo with a -2.7 accuracy change) and calibrated the context-length description in the caption.

Weakness 2: Simplifying Figure 2 and Highlighting High-Level Intuition

We have substantially streamlined Figure 2 (page 5 of the revised manuscript) to improve readability while preserving the high-level design logic:

  • Clarifying the top-level pipeline: The four-stage flow (1. Historical memory -> 2. RecordFolder -> 3. ContextFolder -> 4. Retrieval & answering) now summarizes each stage's input-output transformation in a single phrase.
  • Retaining only core expressions: In (a) RecordFolder, we moved secondary density-peak, iterative novelty, and timestamp aggregation formulas to Section 3.3, keeping only the core entity-semantic affinity and temporal matching expressions supported by visual node illustrations.
  • Highlighting ContextFolder's closed-loop logic: In (b) ContextFolder, we added a prominent Core idea summary box (Planner fact-level questioning, Compressor syntax-guided encoding, and Verifier blind recovery check) with concise bullet points and clear Accept / Fallback branches.

To further substantiate our dual-axis efficiency advantage highlighted in your Strength 1, we also added plug-and-play integrations on SimpleMem, LightMem, and REMem and end-to-end latency analyses (detailed in our responses to Reviewers 33d6 and BjJm). Thank you again for your helpful feedback!

AuthorsOpenReview ↗

Follow-up on Our Rebuttal and Manuscript Revisions

Dear Reviewer CviY,

Thank you again for your time and positive review. As the discussion period is drawing to a close, we wanted to briefly follow up to ensure our manuscript revisions and additional experiments have addressed your suggestions:

  • Simplified Figure 1 & Clarified Metric Scopes (W1): We redesigned Figure 1 (page 2 of the revised manuscript) to remove cluttered numbers, add a clear retrieval-unit legend and failure-zone labels, directly contrast pruning, merging, prompt compression, and MemFolder's dual-axis folding, and clarify metric scopes in the caption (99.0% Mem0 retention at 50% reduction; -40.2% tokens on LoCoMo).
  • Streamlined Figure 2 & High-Level Intuition (W2): We simplified Figure 2 (page 5) by summarizing the top 4-stage pipeline in single phrases, moving secondary formulas to Section 3.3 while retaining core affinity/matching expressions, and adding a prominent Core idea summary box with clear Accept/Fallback branches for ContextFolder.
  • Additional Dual-Axis Validation (Strength 1): We also added controlled plug-and-play experiments on SimpleMem, LightMem, and REMem alongside full offline and online end-to-end latency analyses (detailed in our responses to Reviewers 33d6 and BjJm), further substantiating MemFolder's dual-axis efficiency advantage.

We would be grateful to know if these revisions address your suggestions, and we remain happy to answer any remaining questions.

Best regards, The Authors

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.