acceptodds
Under review as a conference paper at ICLR 2027

NexMem: Efficient Multi-Granular Conversational Memory using N-ary Link Scoring

Abstract

Long-term conversational memory methods must preserve specific facts and connect experiences across sessions. Many existing approaches support this through costly, iterative memory refinement steps. We propose NexMem, which jointly extracts memory records and their indexing metadata in a single pass, without subsequent iterative refinements. Typed n-ary links, called nexus points, organize each memory at two granularities: broad aboutness links capture topical context, while fine-grained event-fact links capture entity–phrase associations. Retrieval first scores conventional memory, entity, and phrase graph nodes against the query, then propagates relevance between nexus points using personalized PageRank within a bounded local neighborhood. Across three long-term conversational-memory benchmarks, NexMem achieves the highest average accuracy among evaluated methods on each benchmark, including a 28-percentage-point advantage over the strongest evaluated baseline on the benchmark with the longest histories. On the benchmark where we profile construction cost, NexMem requires half as many LLM calls as the next-lowest profiled memory system and generates 5-10x fewer construction output tokens than the two strongest memory baselines.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
long term memory, retrieval augmented generation, information retrieval, agent memory, graph

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_cMp6OpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper targets the construction cost of long-term conversational memory. The authors argue that existing systems either refine memory over multiple rounds or pay for hierarchical structure to recover cross-session relations, and that both spend a large number of LLM calls at build time. NEXMEM extracts memory records together with entities, phrases and facts in one pass, then organizes them into two kinds of nexus point. An aboutness nexus point binds one memory to every entity and phrase it mentions, giving a broad topical view; an event-fact nexus point binds the memory to a single entity-phrase pair with arity three, giving a narrow factual view.

Critical strengths
  1. The paper considers memory content, entities, phrases, and cross memory relations in a unified system.
  2. The single pass construction procedure can reduce the language model cost compared with multi stage memory processing pipelines.
Critical weaknesses
  1. The main novelty claim surrounding the n ary nexus point is not sufficiently convincing. The event fact nexus point simply connects a memory node with an entity node and a phrase node. This structure can be decomposed into multiple pairwise edges without losing the shared node, adjacency, or propagation information used by the proposed retrieval procedure. The aboutness nexus point is even closer to a set membership relation between a memory and a collection of entities and phrases. It does not model roles, ordering, directionality, argument structure, or verifiable higher order semantics.
  2. MemGAS already uses memory units and indexing information at different levels to support hierarchical retrieval. Memora separates memory content from a lightweight cue anchor indexing layer, allowing the same fact to be accessed through multiple paths. HyperGraphRAG and related hypergraph retrieval methods have also studied multi node associations and higher order connections.
  3. The single pass construction procedure is primarily an engineering optimization rather than a new memory modeling mechanism.
Questions for authors

The paper reads more like an effective engineering system than a clearly novel memory modeling method.

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

This work investigates long-term conversational memory: agents must recall facts and connect related experiences across many sessions, but existing systems retrieve flat chunks that miss fine-grained cross-session relations, use pairwise graph edges that cannot express the n-ary co-occurrence patterns of dialogue, or buy structure with costly iterative LLM refinement. To address it, the authors proposed NexMem, which extracts memory records with their indexing metadata in one LLM call per segment, organizes each memory into an "aboutness" nexus point plus arity-3 "event-fact" nexus points, and retrieves via hybrid search that seeds a bounded, type-aware personalized PageRank over the nexus graph, using two LLM calls per query. Different from existing work that focuses on flat retrieval or pairwise edges, NexMem scores n-ary links directly. Experimental results verify its effectiveness and superiority: best average accuracy on all three benchmarks (86.62 on LoCoMo vs. 82.86 for the strongest baseline; 88.20 on LongMemEval vs. 87.60; 79.98 on LifeBench vs. 51.07), with the fewest LoCoMo construction calls (272 vs. 544 for the next-lowest).

Critical strengths
  • Best average accuracy on all three benchmarks, including above the full-context reference on LoCoMo (86.62 vs. 85.00), with the best single-hop (92.63) and temporal (85.67) scores there and the best multi-hop score among memory systems (78.72 vs. 74.82) (Table 1).
  • Good methodological hygiene: hyperparameters are fixed across benchmarks with no per-benchmark tuning, and the full extraction, anchor, answer, and judge prompts are provided in the appendix.
Critical weaknesses
  • Efficiency is only profiled on LoCoMo. All call and token numbers come from that one benchmark; there are no construction or retrieval costs for LongMemEval or LifeBench, where histories are far longer and the efficiency pitch matters most, and no wall-clock latency is measured anywhere despite "Efficient" in the title.
  • The retrieval-cost comparison is selective: the text claims several-fold fewer tokens than the graph-based systems, but HippoRAG 2 (Gutiérrez et al., 2025) — also a graph/PPR system — uses fewer input tokens per query than NexMem (2.8K vs. 2.92K), and LightMem fewer still (Table 3a).
  • The ablations do not isolate the claimed contribution: removing event-fact points, Level 1, or Level 2 scoring costs at most 0.91 points, and even the weakest ablation still beats every LoCoMo baseline, so most of the gain may come from the extraction prompt and hybrid search rather than the nexus machinery — yet there is no control condition using the extracted memories with plain retrieval and no nexus points or PageRank (Table 3b).
  • Evaluation confounds are not controlled: a single model (Qwen3.5-27B) serves as memory agent, answer generator, and judge for all methods, with a judge prompt that instructs generous grading; whether baselines also used NexMem's elaborate answer prompt — whose recency-based conflict-resolution rules could selectively inflate exactly the categories driving the LongMemEval margin — is never stated, and baseline embedding models are never specified.
  • Lack thorough research on recent work, missing relevant work discussions like AssoMem.
Questions for authors
  • Why is NexMem best on LoCoMo temporal questions (85.67) yet 15.79 points behind Memora on LongMemEval temporal reasoning? Is this because the anchor-extraction prompt disables the temporal filter for relational-time questions ("before, after, later, earlier…"), which is precisely what that category tests?
  • Was the answer-generation prompt (thinking tags, recency-based conflict resolution) used for all systems or only for NexMem? If only NexMem, how much of the knowledge-update and multi-session margins survives when baselines use the same prompt?
  • How are the "initial retrieved evidence" in the anchor-extraction prompt and the character profile in the answer prompt produced, given that the algorithm extracts anchors from the query alone before any retrieval has run? Neither is defined in the method section.
  • What does "(−) level 1 scoring" mean operationally — if Level 1 is removed, what seeds Level 2? And can you add the missing control: the extracted memories retrieved by hybrid search alone, with no nexus points or PageRank?
Reviewer_SED4OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes NexMem, a conversational memory system that extracts memories and indexing metadata in one LLM call per conversation segment. It combines broad aboutness links and fine-grained event-fact links with hybrid search and local personalized PageRank. Experiments report strong accuracy across three benchmarks, with lower construction costs measured on LoCoMo.

Critical strengths

1.The single-pass memory construction is simple and practical. On LoCoMo, it substantially reduces construction calls and output tokens compared with the closest baselines in accuracy.

2.The method achieves strong results across three benchmarks, especially on LifeBench. Keeping the same hyperparameters across datasets also supports its practical usefulness.

Critical weaknesses
  1. The efficiency evaluation is incomplete. The paper mainly reports LLM calls and token counts on LoCoMo, without measuring end-to-end construction and retrieval time. This makes the practical accuracy–cost trade-off harder to assess, especially for the much longer histories in LifeBench.

  2. The contribution of the nexus structure is not clearly isolated. The ablations only cover LoCoMo, and a hybrid-retrieval baseline using the same extracted memories is missing, making the source of the gains unclear.

Questions for authors
  1. How does NexMem compare with hybrid retrieval over the same extracted memories, using the same embedding model and context budget? A comparison on a representative LifeBench subset would help clarify how much of the improvement comes from the nexus structure.

  2. Could you provide end-to-end construction and per-query retrieval times on LoCoMo and LifeBench, including embedding and graph operations? Algorithm 1 appears to score all nexus points before selecting the local neighborhood. Is this what the implementation does, and how does that step scale with memory size?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.