acceptodds
Under review as a conference paper at ICLR 2027

CorMem: A Brain-Inspired Bipartite Cognitive Map Inducing Latent Relations for Accurate Long-Term Memory Retrieval in LLM Agents

Abstract

Long-term LLM agents require memory systems that can retrieve relations across past experiences while incorporating new information without corrupting stored evidence. Relational organization plays a central role in long-term memory, where experiences are organized through associative and hierarchical structures. Current graph-based memory systems attempt to capture such relational structures. For example, association graphs capture broad semantic similarities but entangle meanings, whereas explicit-relation graphs depend on explicit relation extraction but may omit implicit relations. However, designing the representation for relationships remains challenging, as existing methods struggle to achieve both semantic explicitness and broad relational coverage. Inspired by infants' ability to infer latent relationships from statistical co-occurrence, we ask whether relational structure can be naturally induced. To this end, we propose CorMem, a dual-layer memory that induces latent relations in a cue–fact bipartite cognitive map and forms higher-level semantic gists in a hierarchical schema layer. Cue co-occurrence creates shared-cue paths between facts, and query-conditioned Personalized PageRank (PPR) traverses this structure to activate relevant memories. Schema formation incorporates updates without rewriting stored facts. Experiments on LoCoMo and PersonaMem-v2 show consistent gains over strong baselines. Across all LoCoMo backbones, CorMem achieves the best average F1 and improves gold-session Recall@3 by up to 18.4 points. Retrieved session co-activation also aligns with annotated co-evidence. On PersonaMem-v2, it reaches 70.5% accuracy and significantly outperforms the variant without the schema layer.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
LLM agents, long-term memory, graph-based memory, latent relations, cognitive map, schema formation

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_WLrvOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents CorMem, a long-term memory system for LLM agents that combines a cue–fact bipartite graph with a hierarchical schema layer. The authors report improvements over several baselines on LoCoMo and PersonaMem-v2.

Critical strengths

The paper is well written and generally clear.

Critical weaknesses

1.The structural ablations in the main text change both the graph topology and the types of query seeds, making it difficult to isolate the contribution of the bipartite structure itself. Could the authors clarify whether the bipartite structure still outperforms simpler alternatives when the query information and candidate retrieval conditions are kept as similar as possible? Regarding Table 2, why does the higher recall translate into only a small improvement in answer quality? How much practical benefit does the full structure provide over the simpler structure? 2.The paper does not evaluate whether the hierarchical organization of gists and graph propagation are necessary. Would a comparison between flat retrieval and hierarchical graph-based retrieval demonstrate an independent benefit from the more complex structure? 3.I also have concerns about how the system handles conflicts between new instructions and existing abstractions. The paper aggregates historical evidence through shared cues and hierarchical gists, but the main text does not clearly explain how such conflicts are resolved. In these situations, how does the system ensure that the latest constraint is retrieved and given priority? Are old gists revised, marked as invalid, or assigned lower priority, or does the system simply rely on the answer model to resolve conflicts after seeing the conflicting information? If the new constraint is not included in the final retrieval results, how does the system avoid continuing to use the outdated preference?

Questions for authors

Please see the specific concerns and questions raised in the Weaknesses section.

6 comments

We sincerely thank the reviewer for the careful examination of CorMem’s key mechanisms and the constructive suggestions for strengthening the ablation analysis. Below, we present additional controlled experiments to clarify the contributions of the cue–fact graph and hierarchical schema, and explain how the system handles memory updates.

W1-1. Contribution of the cue–fact structure to retrieval

We thank the reviewer for suggesting controlled comparisons. We examine full versus flat retrieval with the same memory units, cue–fact conversion in HippoRAG 2 (Gutiérrez et al., 2025), and evidence recovery from cue seeds alone.

1. Practical benefits of the full structure over simpler structures

Using the same stored facts and gists, we compare full CorMem with a simpler flat alternative that retrieves these units directly.

MethodLoCoMo F1 (%) ↑LoCoMo BLEU-1 (%) ↑PersonaMem-v2 Accuracy (%) ↑
Flat53.5245.3066.98
Full CorMem55.28†47.00†70.48†

† Significant improvement over Flat (p<.05p<.05).

LoCoMo F1 improves by 1.76 percentage points, with significant gains in both F1 and BLEU-1 (Holm-adjusted p=.0352p=.0352 for both). PersonaMem-v2 accuracy improves by 3.50 percentage points (p=3.71×10−6p=3.71\times10^{-6}). These gains demonstrate the practical benefit of the full structure over the simpler flat alternative, with memory content held fixed.

2. Cue–fact conversion improves HippoRAG 2 by 3.46 points

To test the transferable benefit of the cue–fact structure, we reuse HippoRAG 2’s frozen OpenIE outputs (entity1, relation, entity2) and replace the direct entity1–entity2 connection with entity1–fact–entity2. GPT-4o-mini minimally verbalizes each triple as a fact sentence to express the original relation unambiguously. We retain HippoRAG 2’s models, candidate budget, PPR parameters, and Top-3 budget, while using verbalized fact embeddings and adding a fact-node restart channel. We evaluate the conversion on LoCoMo with GPT-4o-mini.

VariantRecall@3 (%) ↑
Native graph73.93
Cue–fact conversion77.39†

† Holm-adjusted p<.05p<.05 versus the native graph.

Recall improves across all four question categories and all ten memory stores, with an overall gain of 3.46 percentage points (95% CI [2.19, 5.25], Holm-adjusted p=.004p=.004). These gains demonstrate the benefit of introducing the cue–fact representation into HippoRAG 2 using its frozen OpenIE evidence and existing retrieval framework.

3. Cue–fact propagation achieves 79.23% Recall@3 without fact seeds

Under the Table 2 protocol, we reuse the original query-derived seeds on the unchanged graph and zero only the fact-seed weights before renormalization.

VariantRecall@3 (%) ↑
w/o fact seeds79.23
Full CorMem86.83†

† p=.002p=.002 versus w/o fact seeds.

Cue-seeded propagation achieves 79.23% Recall@3, above HippoRAG 2’s 73.93%, showing that supporting evidence remains accessible through cues even without direct query–fact activation. (disabling cue seeds also significantly reduces Recall@3 to 84.08%, p=.004p=.004)

W1-2. Why higher recall yields smaller F1/BLEU-1 gains

Three factors help explain the difference between Recall@3 and F1/BLEU-1 gains in Table 2.

  1. Lexical scoring noise. F1/BLEU-1 emphasize wording overlap, which can obscure answer correctness (Ru et al., 2024). An incorrect answer can therefore outscore a correct but differently phrased one.
  2. Non-proportional changes. Retrieval coverage and generation quality are distinct evaluation targets (Ru et al., 2024). Our grouped results show that greater coverage does not correspond to proportional F1/BLEU-1 gains, particularly at partial recall.
  3. Generation variability. Generated answers introduce sampling variability into evaluation (Zhang et al., 2026). We reduce its impact by averaging three generations per question under a shared answer prompt.

The following example illustrates the first point.

Question
What kind of stress was Sam dealing with in October 2023?

Reference answer
Work-related stress.

MethodGenerated answerRecall@3 (%)F1 (%)BLEU-1 (%)
CorMem (correct)“Work stress, family pressure regarding health.”100.0025.0016.67
Flat (incorrect)“Health-related stress.”0.0050.0050.00

CorMem identifies work stress and adds dialogue-supported detail, yet Flat’s incorrect answer receives higher scores.

For the second point, we group Full CorMem’s questions by gold-session recall. The two higher partial-recall groups have nearly identical answer scores, while complete recall corresponds to substantially higher scores.

Gold-session Recall@3 RRF1 (%)BLEU-1 (%)
R=0R=022.7816.97
0<R≤1/30<R\leq1/323.2118.11
1/3<R≤2/31/3<R\leq2/333.0222.83
2/3<R<12/3<R<132.2822.66
R=1R=162.5154.62

Although F1/BLEU-1 depend on reference answers and model-generated wording, they remain widely used metrics for evaluating memory systems. We retain them for comparable end-to-end evaluation, alongside Recall@3 as a core retrieval metric that directly measures evidence coverage independently of answer generation.

W2. Benefits of hierarchical gist organization and graph propagation

We sincerely thank the reviewer for suggesting these comparisons. W1-1 shows that the full structure outperforms flat retrieval with the same memory units (LoCoMo F1 53.52% → 55.28%, Holm-adjusted p=.0352p=.0352, and PersonaMem-v2 accuracy 66.98% → 70.48%, p<.001p<.001). Below, we separately examine the contributions of graph propagation and gist hierarchy.

1. Graph propagation adds 3.12 Recall@3 points with fixed seeds

To isolate the contribution of graph propagation, we disable fact–cue–fact traversal and activate sessions directly from the seeds.

We follow the Table 2 protocol on LoCoMo, keeping extracted memory, queries, and initial seed activations identical.

VariantRecall@3 (%) ↑
w/o fact–cue propagation83.71
Full CorMem86.83†

† p<.05p<.05 versus w/o fact–cue propagation.

Propagation improves Recall@3 by 3.12 percentage points (p=.004p=.004). With extraction and initial activation fixed, this gain demonstrates the benefit of traversing shared-cue paths to retrieve supporting evidence.

2. Gist hierarchy improves accuracy by 4.34 points

To examine hierarchical gist organization, we flatten the gists into a single layer by removing gist–gist edges and connecting higher-level gists directly to their supporting facts.

Queries, initial seed activations, and the facts retrieved from the cognitive map remain identical. We focus on the 311 queries whose retrieved gist units change. The remaining queries retrieve the same memory units, so this diagnostic focuses on changes in retrieved content while reducing noise from answer generation. Queries are selected solely by retrieval differences, independently of answer correctness.

Gist graphAccuracy (%) ↑
Flattened gist graph60.29
Original hierarchy64.63†

† p<.05p<.05 versus the flattened gist graph.

On this diagnostic subset, retaining the hierarchy improves accuracy by 4.34 percentage points (95% cluster-bootstrap CI [0.81, 7.99], p=.0352p=.0352). This supports the hierarchy’s retrieval role, allowing higher-level gists to aggregate PPR activation from multiple descendants and guide the selection of relevant abstractions.

Reviewer_96JvOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

The paper investigates memory in long-term LLM agents. To address the overlooking of implicit relations, the authors propose CORMEM, a memory structure including shared-cue paths between facts and query-conditioned Personalized PageRank (PPR). Experiments on LoCoMo and PersonaMem-v2 show improvement.

Critical strengths

The paper proposed a novel memory structure and update mechanism, experimentally demonstrating improvements on select datasets.

Critical weaknesses

The innovations include a memory structure and an update mechanism, but the paper does not propose or address a previously overlooked question. The reasons behind the inability and utility remain unexplored, resulting in limited takeaways for future work.

Questions for authors

More practical examples should be added to help readers understand the challenges and the reasons behind the proposed methods.

The relationship with existing work is unclear. Intuitively, associative capabilities in LLMs have been widely investigated in existing work. What additional capabilities does the proposed work offer beyond existing work, such as MM-OPERA[1] and The Labyrinth of Links[2]?

[1]MM-OPERA: Benchmarking Open-ended Association Reasoning for Large Vision-Language Models

[2] The Labyrinth of Links: Navigating the Associative Maze of Multi-modal LLMs

There are incorrect citations in the main paper. For example, LoCoMo and PersonaMem-v2 should be cited when they first appear in the sentence “Experiments on LoCoMo and PersonaMem-v2 support the proposed memory structure.”

The conclusion only shows improvements on selected datasets, without providing key takeaways for future work or potential explanations for why previous methods failed and how the proposed methods solved these problems.

How is the quality of text processed by LLM tools ensured?

4 comments

We sincerely thank the reviewer for the thoughtful review and for recognizing CorMem’s memory structure, update mechanism, and empirical benefits. We add practical examples, FullText comparisons, and seed-perturbation analyses, and clarify extraction controls.

W1 & Q1. Research motivation and key challenges

Existing memory structures can hinder retrieval of related experiences, while triple extraction may omit latent relations (Xu et al., 2025; Gutiérrez et al., 2025).

We propose a cue–fact graph that preserves source-grounded facts and induces latent relations through shared-cue co-occurrence, without predicting pairwise relations. Graph propagation then enables related experiences to be retrieved together.

Example 1. Local associations can obscure relevant connections

This example shows how local associations can overlook relevant evidence even when both supporting memories are stored.

Question — Where did Audrey get Pixie from?
Reference answer — A breeder.

Earlier dialogue — “We adopted another puppy called Pixie.”

Later dialogue — “I got lucky finding a breeder nearby that has the dogs I wanted. Yeah places that allows dogs are really hard to find…”

In A-MEM, all five outgoing links from the breeder memory lead to housing-related memories, while the question concerns the puppy’s origin.

CorMem connects the experiences through their shared Audrey cue.

Pixie adoption fact → Audrey cue → breeder fact

This path brings both facts into the answer context, and CorMem correctly answers “a breeder.”

Example 2. Triple extraction can omit question-relevant information

This example shows how triple extraction retains actions but omits the motivations needed to answer the question.

Question — What has Maria done to feel closer to her faith?
Reference answer — Bought a cross necklace and joined a church.

Dialogue excerptHippoRAG 2 tripleOmitted information
“I bought a cross necklace to feel closer to my faith…”["Maria", "bought", "cross necklace"]Motivation for the purchase
“Just yesterday I joined a nearby church. I wanted to feel closer to a community and my faith.”["Maria", "joined a church", "Yesterday"]Motivation for joining

CorMem preserves it within the facts.

“Maria bought a cross necklace to feel closer to her faith, which has made her happy.”

“Maria joined a nearby church yesterday to feel closer to a community and her faith, describing the experience as really great so far.”

The shared faith cue connects these separately stored experiences.

Necklace-purchase fact ← faith cue → Church-joining fact

CorMem retrieves both experiences and correctly includes both actions in its answer. Their connection follows from a shared cue, without requiring an explicitly extracted relation between buying the necklace and joining the church.

Q2. LLM associative reasoning and long-term memory

We thank the reviewer and will cite both studies. MM-OPERA evaluates models’ associative reasoning (Huang et al., 2025). The Labyrinth of Links retains inferred associations as short-term memory for subsequent reasoning (Li et al., 2025).

These studies examine how models form and reuse associations. Two findings explain why reasoning over past experiences also requires effective memory organization.

  1. LongMemEval shows that LLMs answer better from relevant evidence sessions than from complete histories, highlighting the need to retrieve the right experiences (Wu et al., 2025).
  2. HippoRAG 2 highlights the need to retrieve connected facts together, which independent vector retrieval struggles to achieve (Gutiérrez et al., 2025).

Together, these findings motivate organizing memories so that related evidence can be retrieved jointly. CorMem addresses this problem by preserving experiences as facts and inducing cue-mediated latent relations through shared-cue co-occurrence.

Table 1a compares CorMem’s retrieval of three raw dialogue sessions with FullText’s complete dialogue history, using the same answering models and prompt.

LoCoMo average scores (%), session-retrieval evaluation in Table 1a.

BackboneF1 (%) ↑ (FullText → CorMem)BLEU-1 (%) ↑ (FullText → CorMem)
GPT-4o-mini37.67 → 55.8429.93 → 47.98
Qwen3.5-27B54.60 → 62.2146.87 → 54.01
DeepSeek-v4-Flash49.19 → 61.1341.35 → 53.11
MethodMemory tokens/query ↓
FullText27,083.5
CorMem559.1

CorMem uses the Top-10 setting (Table 4).

CorMem improves F1 and BLEU-1 across all three backbones, while the Top-10 comparison shows substantially lower context requirements. These results support memory organization as a practical complement to the model’s associative reasoning.

Q3. Key takeaways and future directions

We thank the reviewer and apologize for the brief discussion. We will correct presentation issues and clarify the conclusion. Below, we summarize two takeaways on relation formation and retrieval, and outline future directions.

1. Cue co-occurrence induces latent relations.

CorMem links each fact to its cues. As experiences accumulate, shared cues connect facts without requiring explicit pairwise relation extraction (W1 & Q1).

2. The induced structure supports retrieval beyond direct fact matching.

We add a cue-only initialization test under the Table 2 protocol, keeping the graph unchanged and zeroing only fact-seed weights before renormalization.

InitializationRecall@3 (%) ↑
Cue seeds only79.23
Full initialization86.83

Shared-cue paths therefore provide access to supporting evidence even without direct query–fact activation. In separate perturbation tests, Recall@3 remains 70.46% when 80% of initial seeds are randomly deactivated and 78.60% when 40% are replaced with random nodes of the same type.

The broader implication is that accumulated relations can support retrieval from incomplete or noisy cues. Future work will explore visual and auditory features as cues, investigating whether their co-occurrence can similarly organize experiences and provide complementary paths for memory retrieval.

Q4. Quality of LLM-processed memory

We thank the reviewer for this question. Below, we explain how CorMem preserves experience content during extraction and updates, supported by explicit prompts and validation controls.

Existing work addresses memory quality through source grounding, consistency checks during updates, and downstream answer-quality evaluation (Park et al., 2023; Chhikara et al., 2025; Fang et al., 2026).

CorMem focuses on preserving evidence during extraction and updates, retaining contextual meaning within facts and managing evolving preferences through a separate gist layer.

1. The cue–fact structure preserves context.

CorMem retains contextual information within facts and connects them through shared cues, without extracting pairwise relations. This preserves attribution, motivations, and constraints that triple extraction can omit, as illustrated in W1 & Q1.

2. Separate fact storage from memory updates.

CorMem keeps committed facts unchanged and uses gists to represent evolving preferences, constraints, and abstractions. This allows new information to guide answers while preserving earlier evidence. Table 3 supports this design over rewriting stored facts.

Update strategyDiagnostic LoCoMo F1 (%) ↑PersonaMem-v2 Accuracy (%) ↑
Rewrite51.5669.10
Separate schema formation54.3170.48

Our entity-first prompts require extraction from the supplied dialogue, preserving supported details without adding outside knowledge. Source pointers, timestamps, structured-output validation, and gist-coherence audits provide further controls (Section 3.2 and Appendices C, D, and K).

Reviewer_wneqOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

In this work's proposed method, an LLM extracts atomic facts from each dialogue session, and each fact is linked to entity, topic, and episode cues in a bipartite cue-fact graph. Facts are never rewritten. A separate schema layer groups facts into hierarchical gists. Retrieval runs PPR from reranked fact and cue seeds, with a second pass over gists. On LoCoMo and PersonaMem-v2, CoreMem outperforms baselines, and this work adds ablations and a relational alignment analysis.

Critical strengths
  1. CorMem retrieves gold sessions much more reliably than HippoRAG 2 across all three backbones. Both methods return three raw sessions and use the same answer prompt, making this a useful comparison. The authors also report cluster-bootstrap confidence intervals and Holm-corrected tests, which shows overall gains are solid.

  2. The schema layer applies corrections and forgetting requests as answer-time constraints. On PersonaMem-v2, removing it costs more than half of CorMem's lead over Mem0, a statistically significant drop. This addresses a real weakness of memories that overwrite their own evidence.

Critical weaknesses
  1. The cognitive map links facts to their entities, topics, and source sessions, and retrieval uses PPR. AriGraph already links each observation to the triplets extracted from it and ranks observations by the retrieved triplets they contain. The cue-mediated latent relations are ordinary two-hop paths through shared nodes. The schema layer's cluster and summarize hierarchy resembles [1]. The main additions seem to be typed cue-fact restriction, the constraint gists and the combination, so the conceptual novelty is modest. Besides, the set of baselines seems to miss several work in the same methodology direction.

  2. The paper credits the controlled gain to graph traversal. Yet its wo graph control is plain dense session ranking, so its large drop does not isolate propagation. CorMem also differs from HippoRAG 2 in extraction and in picking PPR seeds with BM25 and a cross-encoder reranker. In the GPT-4o-mini ablation, the fact-only variant retains most of the Recall@3 margin over HippoRAG 2, but it still uses cue-derived fact–fact edges. This leaves it unclear how much of the gain comes from extraction and seed selection, and how much comes from the cue-fact structure.

[1] Sarthi, Parth, et al. "Raptor: Recursive abstractive processing for tree-organized retrieval." International Conference on Learning Representations. Vol. 2024. 2024.

Questions for authors

None

3 comments

We sincerely thank the reviewer for the thorough and constructive review, and for recognizing the retrieval performance of our cue–fact memory and the value of separating schema formation from fact storage. The reviewer’s concrete suggestions have helped us clarify CorMem’s relation to prior work and strengthen the analysis of its structural contribution. We expand comparisons with recent memory systems and present the cue–fact transfer to HippoRAG 2 alongside additional controlled ablations of propagation and seed initialization. We address each point below.

W1. Relation to prior work and baseline comparisons

We thank the reviewer for highlighting these connections. CorMem shares the goal of structured memory with AriGraph (Anokhin et al., 2025), HippoRAG 2 (Gutiérrez et al., 2025), and MAGMA (D. Jiang et al., 2026). The distinction lies in how relations are represented and used. AriGraph links observations to extracted triples and ranks observations by retrieved-triple coverage. CorMem induces cue-mediated latent relations through shared cues over source-grounded facts, without predicting pairwise relations.

We also evaluate AriGraph under the same Top-3 session retrieval setting on LoCoMo with GPT-4o-mini.

MethodRecall@3 (%) ↑
HippoRAG 273.93
AriGraph62.43
CorMem86.83

This distinction matters when triple extraction omits part of an experience’s meaning, as illustrated by our HippoRAG 2 example.

What has Maria done to feel closer to her faith?

Reference answer
Bought a cross necklace and joined a church.

Original dialogueExtracted HippoRAG 2 tripleOmitted information
“I bought a cross necklace to feel closer to my faith…”["Maria", "bought", "cross necklace"]Purpose of the purchase
“I joined a nearby church. I wanted to feel closer to a community and my faith.”["Maria", "joined a church", "Yesterday"]Purpose of joining the church

The triples preserve the actions but lose their connection to faith.

CorMem stores the following two facts.

“Maria bought a cross necklace to feel closer to her faith, which has made her happy.”

“Maria joined a nearby church yesterday to feel closer to a community and her faith, describing the experience as really great so far.”

Under matched Top-3 session retrieval, CorMem answers with both actions, while HippoRAG 2 answers with community work. Their three-run mean F1 scores are 66.67% and 6.50%, respectively.

Each fact–cue–fact path represents a shared-cue association. As memory accumulates, query-conditioned PPR aggregates activation across these paths, allowing multiple associations to contribute to evidence ranking. Figure 4 supports this task-aligned relational organization. CorMem’s session co-activation has the highest mean alignment with annotated co-evidence (ρ = .248, versus .218 for LightMem and .161 for HippoRAG 2).

Controlled tests of the structural contribution are presented in W2.

RAPTOR builds hierarchical summaries for document retrieval (Sarthi et al., 2024), whereas CorMem’s separate schema layer additionally encodes preferences and update constraints that govern how retrieved evidence is used, without rewriting stored facts.

We also compare CorMem with recent baselines using their reported LoCoMo F1 scores (GPT-4o-mini, Categories 1–4).

MethodAverage F1 (%) ↑
CorMem52.79
AnchorMem (Shen et al., 2026)49.87
Synapse (H. Jiang et al., 2026)40.50
HiGMem (Cao et al., 2026)40.46
MAGMA (D. Jiang et al., 2026)46.70

External results follow each study’s evaluation protocol. HiGMem’s average uses question-count weighting.

We thank the reviewer for these suggestions and will expand the main-text discussion of both works (AriGraph and RAPTOR), with corresponding citations.

W2. Analysis of component contributions

We thank the reviewer for suggesting further controls to distinguish the contributions of memory structure, graph propagation, and seed initialization.

1. Cue–fact conversion improves HippoRAG 2 by 3.46 points

To test the transferable benefit of the cue–fact structure, we reuse HippoRAG 2’s frozen OpenIE outputs (entity1, relation, entity2) (Gutiérrez et al., 2025) and replace the direct entity1–entity2 connection with entity1–fact–entity2. GPT-4o-mini minimally verbalizes each triple as a fact sentence to express the original relation unambiguously.

VariantRecall@3 (%) ↑
Native graph73.93
Cue–fact conversion77.39†

† Holm-adjusted p<.05p<.05 versus the native graph.

Recall improves across all four question categories and all ten memory stores, with an overall gain of 3.46 percentage points (95% CI [2.19, 5.25], Holm-adjusted p=.004p=.004). These gains demonstrate the benefit of introducing the cue–fact representation into HippoRAG 2 using its frozen OpenIE evidence and existing retrieval framework.

On LoCoMo (GPT-4o-mini), we retain HippoRAG 2’s models, candidate and Top-3 budgets, and PPR parameters, while using verbalized fact embeddings and adding a fact-node restart channel.

2. Propagation and seed ablations

To isolate the contributions of graph traversal and seed selection, we extend the ablations in Table 2.

VariantRecall@3 (%) ↑
w/o fact–cue propagation83.71
w/o fact seeds79.23
w/o cue seeds84.08
Full CorMem86.83†

† p<.05p<.05 versus each ablated variant.

Propagation improves Recall@3 by 3.12 percentage points (p=.004p=.004), demonstrating a retrieval benefit from graph traversal itself. Fact seeds contribute 7.60 points (p=.002p=.002), and cue seeds contribute 2.75 points (p=.004p=.004). Even without direct semantic activation of facts, the cue–fact structure still recovers associated evidence through shared cues (79.23% Recall@3).

For each query, controls share the same memory and initial seed activations. w/o fact–cue propagation blocks fact–cue–fact traversal and activates sessions directly from seeds. The seed ablations retain the full graph and zero only the corresponding seed weights.

Thanks for the comparison and the ablation. This gives the clean control I was asking for and shows that graph propagation adds a significant but modest share of the gain over HippoRAG 2, with extraction and seed selection contributing to most of it. This is consistent with my assessment, so I keep my current score.

Best,

Reviewer wneq

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.