acceptodds
Under review as a conference paper at ICLR 2027

SAGE: A Self-Evolving Agentic Graph-Memory Engine for Structure-Aware Associative Memory

Abstract

Long-term memory is becoming a central bottleneck for language agents. Existing RAG and GraphRAG systems largely treat memory graphs as static retrieval middleware, which limits their ability to recover complete evidence chains from partial cues, exploit reusable graph-structural roles, and improve the memory itself through downstream feedback. We introduce SAGE, a Self-evolving Agentic Graph-memory Engine that models graph memory as a dynamic long-term memory substrate. SAGE couples two roles: a memory writer that incrementally constructs structured graph memory from interaction histories, and a Graph Foundation Model-based memory reader to perform retrieval and provide feedback to the memory writer. We provide rigorous theoretical analyses supporting the effectiveness of carefully designed architectural components and the framework. Across multi-hop QA, open-domain retrieval, domain-specific review QA, and long-term agent-memory benchmarks, SAGE improves evidence recovery, answer grounding, and retrieval efficiency: after two self-evolution rounds, it achieves the best average rank on multi-hop QA; in zero-shot open-domain transfer, it reaches EM/F1 of 46.20/54.21 and 53.00/66.06 on NQ and PopQA. Most importantly, when used as a replacement memory-access backend for existing memory systems, SAGE improves LongMemEval overall accuracy by 7.4 to 9.6 absolute points over the native MemOS, Memobase, and Zep backends after one evolution round, while paired HaluMem experiments improve most operation-level memory metrics and reduce QA hallucination. These results suggest that self-evolving, structure-aware graph memory is a promising way for robust long-horizon language agents. Our code is available https://anonymous.4open.science/r/Unified-Representation-A9D9/here.

Primary area:
foundation or frontier models, including LLMs
Keywords:
long-term agent memory, self-evolving memory, Graph Foundation Models, agentic memory systems

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_M8vcOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes SAGE, which models long-term memory as a learnable system for graph-based memory writing and reading.

Critical strengths

I appreciate the idea of linking the quality of memory writing to its subsequent use. This is better aligned with the needs of long-term memory systems than optimizing information extraction or retrieval in isolation.

Critical weaknesses

1.The writer already has access to the current query when constructing the graph. There are two different tasks here: organizing historical information first and subsequently answering questions that were not known in advance, and organizing historical information specifically for a query that is already known. Does graph construction at test time depend on the current query? Under this design, is the same graph reused across multiple queries? Can an improvement in the writing policy benefit previously unseen queries? The authors need to evaluate these aspects to demonstrate the generality and effectiveness of the memory, as well as the method’s potential applicability to real-world scenarios. 2.The paper presents a relatively comprehensive system design, but the main text does not clearly focus its core contribution. I suggest that the authors clarify the new mechanisms and their necessity in relation to the most closely related graph foundation model-based retrieval methods and trainable memory-writing approaches. For a framework that repeatedly emphasizes self-evolution, what exactly is distinctive about the writer and reader? These are common components in memory systems, and I would like the authors to clarify this point. For example, where do the core additional capabilities come from compared with existing GFM-based retrievers?

Questions for authors

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

Reviewer_P2ZnOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This work presents SAGE, a self-evolving graph-memory system for long-horizon language agents. A policy-based memory writer incrementally turns interaction fragments into entities, relations, aliases, and source anchors; a Graph Foundation Model reader performs query planning, soft entity addressing, structurally gated propagation, and context/schema calibration to retrieve documents and subgraphs. Retrieval and answer feedback then reward the writer, and the writer and reader are updated alternately. Experiments cover multi-hop QA, open-domain transfer, AmazonQA, LongMemEval, and HaluMem; reported gains include a better average QA rank after two evolution rounds and improvements when SAGE replaces the retrieval backend of several memory systems.

Critical strengths
  • The writer-reader loop directly optimizes memory writing for downstream use. Instead of treating a graph as a fixed index, SAGE trains the writer with reader-derived rewards for evidence precision, evidence recall, deducibility, answer quality, repetition, and format. This gives the memory-writing policy a concrete objective tied to retrieval utility rather than generic graph construction.
Critical weaknesses
  • The claimed self-evolution still depends heavily on labeled offline supervision. Writer rewards use gold supporting documents, gold answers, an answer-deducibility judge, and answer-alias F1; Algorithm 1 also defines evolution over a training set. Thus, the method is best viewed as supervised writer-reader co-training, not label-free online self-improvement. The paper does not demonstrate the same evolution quality when only unlabeled interaction histories or delayed user feedback are available.

  • Deployment cost and result stability are not yet adequately resolved. The audited pipeline involves large-scale graph construction, multiple LLM calls, GFM pretraining, and alternating writer-reader updates; rebuilding the raw training graphs is estimated at millions of LM calls even before considering evolution rounds. The best fine-tuning compute trace is unavailable, and main comparisons lack repeated runs or confidence intervals. These issues make it difficult to assess practicality and the stability of the reported gains.

Questions for authors
  • In each self-evolution round, exactly which split and labels are used for writer rewards and reader updates? Can the method evolve solely from unlabeled interactions, and how much does performance degrade without gold supporting documents or answers?

  • What is the incremental cost when new memories arrive after deployment? Can the writer and reader be updated without rebuilding the graph or re-running expensive pretraining, and how stable are earlier memories under such updates?

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

The paper introduces SAGE, a graph-memory framework that jointly optimizes how memories are constructed and retrieved. An LLM writer generates entity-relation triples with links to source documents and is trained using retrieval and answer-related rewards. The reader combines structured query planning, global soft addressing, structural gating, and context–schema representations to retrieve supporting documents. Training alternates between improving the writer against a frozen reader and adapting the reader to the resulting graphs. Experiments cover multi-hop question answering, restricted-corpus transfer, review-based question answering, and integration into existing memory systems. The paper reports improved overall LongMemEval accuracy across three matched host systems and provides component and training ablations.

Critical strengths

The paper addresses a relevant problem: memory construction should account for how the stored information will subsequently be retrieved. The paired comparisons within MemOS, Memobase, and Zep are useful because they aim to isolate the effect of replacing the retrieval backend. Table 2 reports improved overall LongMemEval accuracy for all three hosts, suggesting practical value for the proposed component.

The ablation suite is substantial. Table 26 includes a one-shot training control, an equal-compute uncoupled control, a parameter-matched context–schema ablation, and a same-architecture control without pretraining. Full SAGE improves over the equal-compute uncoupled control by 2.2, 3.9, and 3.1 F1 points on the three multi-hop datasets. These comparisons provide meaningful evidence for the design. I also appreciate the appendix’s disclosure of deployment limitations and additional computational costs.

Critical weaknesses

A key theoretical lower bound does not follow from the stated assumptions. In Appendix B, Eq. (22) provides only a coordinate-wise upper bound on the next-layer activation. Lemma B.6 nevertheless derives a lower bound on evidence activation in Eq. (28). Consider two nodes, with the first in the evidence region, an identity propagation operator, zero perturbation, previous activation (1, 1), and next activation (0.25, 1). Eq. (22) holds, and the evidence-retention coefficient can be 1, but Eq. (28) would require 0.25 to be at least 1. Theorem B.7 can still hold conditionally on its assumed recurrences, but the argument connecting those recurrences to the reader is incomplete. This weakens the theoretical justification of the method and is a main reason for my rating.

The memory experiments do not clearly establish the scope of the self-evolution claim. Algorithm 1 uses queries, gold supporting documents, and answers during training. This is a legitimate supervised setup, but the manuscript needs a more explicit account of the training, validation, and evaluation separation for the results in Tables 2 and 17. Appendix S also states that normal MemOS writes do not synchronize new memories into the prebuilt SAGE index. This limits what the reported implementation demonstrates about continuously updated online memory. My concern is insufficient specification of the evaluation protocol. I am not claiming that data leakage occurred.

The incremental benefits of evolution need stronger statistical support and a clearer discussion of tradeoffs. In Table 2, one evolution round adds 1.4, 0.5, and 2.6 percentage points beyond the initial SAGE substitution. These increments should be distinguished from the larger gains over the native backends and supported by paired uncertainty estimates or repeated runs. Improvements are also not uniform: knowledge-update accuracy remains below the native system for MemOS and Memobase, and HotpotQA EM falls from 60.4 after one round to 56.1 after two rounds. Table 26 is helpful, but the exact resource accounting behind its equal-compute control is insufficiently specified.

Questions for authors

What additional condition establishes the evidence-activation lower bound in Eq. (28), and why should that condition hold for the implemented reader? Can the authors repair Lemma B.6 and clarify which subsequent guarantees remain conditional?

For the LongMemEval and HaluMem results, what are the precise training, validation, and evaluation partitions? When does the writer see the query, which labels are available during evolution, and are constructed memories evaluated on previously unseen questions?

Which operations invoke SAGE within each host system? How are new memories reflected in the prebuilt MemOS index? Also, if upstream extraction is held fixed, through which mechanism does the retrieval substitution change the HaluMem extraction metrics?

What resources are matched in Table 26’s equal-compute control, including writer rollouts, LM tokens, reader optimization steps, and graph reconstruction? Can the authors provide uncertainty estimates for the existing evolution comparisons and explain the knowledge-update and EM regressions?

Can the authors provide a matched comparison of complete online latency and LM usage, using the same measurement boundary for SAGE and the relevant baselines, and replace the sum-of-medians argument with directly measured total latencies?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.