PaperlayerView exact revision on arXiv

Source-linked AI summary

EdgeMem: LLM-Free Agent Memory Construction and Retrieval via Evidence-Preserving Multi-Anchor Hypergraph

Zeyang Cui, Jiannong Cao, Zhiyuan Wen, Bo Yuan, Junlan Feng, Shengyuan Chen

arXiv:2609.05553v1cs.AIcs.MA

TL;DR

Existing agent-memory methods can incur repeated LLM cost and discard answer-bearing details when organizing growing multi-session histories. EdgeMem preserves original turns in a multi-anchor hypergraph with LLM-free construction and retrieval, then uses one reader call to answer from retrieved evidence. Across LoCoMo and LongMemEval-S, it delivers strong answer quality and session-level retrieval while using zero generative-LLM tokens for memory construction and retrieval.

  • Problem

    Growing multi-session histories make usable memory and relevant evidence retrieval increasingly challenging, while LLM-based memory processing incurs repeated cost and may discard details.

  • Method

    EdgeMem uses lightweight local annotations to organize original interaction turns in Time, Co-occurrence, and Episode Sub-Hypergraphs, retrieving a source-grounded evidence pack for one reader call.

  • Results

    EdgeMem achieves strong answer quality on LoCoMo and the best session-level retrieval scores on LongMemEval-S, while construction and retrieval use zero generative-LLM tokens.

  • Takeaways & Limitations

    Preserving and organizing source turns provides an efficient, traceable foundation for agent memory without generative memory management.

  • Takeaways & Limitations

    Real-world deployments require consent-aligned retention, access-control, and deletion policies, which are outside the evaluation scope.

Abstract

from arXiv · show

Agent memory allows LLM agents to use earlier interactions when answering new queries. Existing methods often compress interaction histories into summaries or other LLM-generated representations. Repeated generation adds cost and can discard answer-bearing details before the system knows what a future query will require. We propose EdgeMem, an agent-memory method built around a simple principle: preserve original interaction turns and organize them through complementary content, temporal, and episodic cues. EdgeMem realizes this principle with a multi-anchor hypergraph constructed by lightweight local processing. Retrieval directly returns source evidence and reserves LLM use for final answer generation, combining structured access to multi-session histories with faithful retention of the original conversation. Experiments on LoCoMo and LongMemEval-S show strong retrieval and memory-grounded question answering; on LoCoMo, EdgeMem achieves the highest strict-judge score among seven reproduced systems under a shared prompt (61.01 versus 58.70), while construction and retrieval require no generative-LLM calls. Overall, EdgeMem shows that preserving and organizing source evidence provides an effective and efficient foundation for agent memory without generative memory management.

1 Introduction

Agent memory must retrieve useful evidence from growing multi-session histories without losing answer-bearing details or repeatedly paying generative-LLM costs. EdgeMem preserves original turns and organizes them with complementary anchors, achieving strong benchmark performance with LLM-free construction and retrieval.

  • Motivation: Growing multi-session histories make usable memory and query-relevant evidence increasingly difficult to maintain and retrieve.Feeding the full history to every query progressively increases input-token demands and can exceed the model context window.
  • Motivation: Existing memory methods use chunk retrieval or derived representations, but LLM-based construction, updating, and retrieval introduce substantial token costs and evidence-fidelity concerns.Derived summaries or compressed representations may discard answer-bearing details before future queries reveal what matters.
  • Method: EdgeMem retains every interaction turn as atomic evidence and associates it with content, time, and episode anchors in a multi-anchor hypergraph.Its three components are the Time, Co-occurrence, and Episode Sub-Hypergraphs.
  • Method: Deterministic query routing activates matching hyperedges, uses episode context as the core, and adds co-occurrence and time evidence within a fixed budget before one reader call.Local NLP tools and transcript metadata handle annotation and routing; only the final reader invokes an LLM.
  • Results: 61.01 versus 58.70 under LoCoMo’s strict judge was the highest score among seven reproduced systems under a shared prompt.On LongMemEval-S, EdgeMem achieved 81.49 Recall@3 and 90.49 NDCG@3, with an answer score of 60.00 versus 60.20 while using 36% as many tokens per question.

2 Related Work

Prior agent-memory research combines generated abstraction, graph organization, and direct source retrieval, but these strengths are usually separated. EdgeMem addresses this gap by preserving source turns while exposing temporal, episodic, and content-based structure without generative memory management.

  • LLM-written conversational memory: Prior work includes LLM-written summaries, hierarchical stores, extract-and-merge policies, recursive summaries, and compressed topical units.HyperMem further organizes LLM-derived topics, episodes, and facts in a hierarchical hypergraph.
  • Graph-organized memory and retrieval: Graph-organized methods support traversal or diffusion through knowledge graphs, summary graphs, event graphs, multi-granularity graphs, and non-generative document-retrieval graphs.These methods vary in whether graph construction or navigation depends on an LLM.
  • Source-level retrieval and episodic organization: Source-level retrieval grounds answers in selected passages, while BM25 ranks stored text without generation.Flat dialogue retrieval treats turns independently and leaves shared events, episode membership, and time implicit.
  • Source-level retrieval and episodic organization: Episodic-memory accounts motivate preserving source records while exposing content and spatiotemporal context as complementary retrieval cues.This perspective connects conversational recall to shared events, episode membership, and time.
  • Research gap and EdgeMem: EdgeMem combines source preservation, graph association, and direct retrieval through LLM-free construction and retrieval over multiple contextual anchors.The method is introduced as a response to the separation between generation-based memory and flat source retrieval.

3 EdgeMem: LLM-Free Structured Memory Management

EdgeMem preserves source turns and organizes them in a multi-anchor hypergraph using content, time, and episode cues. Deterministic retrieval builds an evidence pack from complementary channels, while a single reader call generates the grounded response.

  • 3.1 Problem Statement: EdgeMem stores each source turn verbatim with entity, lexical, session, and date annotations, making records addressable by content, time, and episode.Local NLP tools and transcript metadata provide the anchor keys without generative-LLM calls.
  • 3.3 Multi-anchor Hypergraph Construction: The multi-anchor hypergraph uses typed time, co-occurrence, and episode hyperedges to connect turns through complementary access cues.Each anchor key induces a hub containing source turns that share that key, while each source record is stored once.
  • 3.4 Memory Retrieval: Query processing extracts date, entity, and lexical cues and routes them through deterministic Time, Co-occurrence, and Episode channels.The Episode channel combines turn- and episode-level lexical relevance; Co-occurrence links recurring content across sessions; Time activates for explicit date spans.
  • 3.4 Memory Retrieval: The Episode channel supplies the evidence core, while Co-occurrence and activated Time evidence fill remaining positions under fixed retrieval budgets.With k=14, n_co=4, n_time=2, and κ=18, the final positions vary according to whether the Date-span Gate is closed or open.
  • 3.5 Generated Answer: The serialized evidence pack retains speaker, date, and unchanged source text before one reader call composes the grounded answer.The reader is the pipeline’s single generative stage and runs at temperature zero with a maximum of 128 output tokens.

4 Experiments

The experiments test EdgeMem’s answer quality, efficiency, evidence preservation, anchor contributions, and transfer across benchmarks. Under shared answer-style controls, EdgeMem combines strong or leading performance with low token and time costs, while evidence-only answering improves unsupported-premise reliability at an answerable-split recall cost.

  • Question-Answering Accuracy (Q1): Removing generative inference from memory construction and retrieval produced no measurable answer-quality cost under the reported comparison.
  • Memory Management Efficiency (Q2): EdgeMem had the lowest token and time costs: roughly 1,000 tokens per question, a 1.8-minute build, and about 1.3 seconds per question.The two closest systems in accuracy used 19,745 and 41,348 tokens per question, while metered builds required 17 minutes to 10.1 hours of API time.
  • Original-Evidence Preservation (Q3): 92.57% of unsupported-premise questions were correctly rejected, exceeding the strongest baseline’s 84.68% under the same refusal instruction.On the answerable split, the evidence-only rule caused abstention on 28.0% of 1,540 questions, with about eight judge points recoverable under an unconstrained prompt.
  • Anchor Ablation Study (Q4): Episode, co-occurrence, and time anchors contributed +8.44, +5.65, and +0.45 judge points over flat BM25, respectively.The episode–co-occurrence overlap makes marginal attribution depend on insertion order, while the final destination remains consistent.
  • Cross-Benchmark Generalization (Q5): 81.49 Recall@3 and 90.49 NDCG@3 were achieved on LongMemEval-S without retuning, while strict-judge accuracy reached 60.00 versus 60.20 for the best transcribed system.The unchanged chain used roughly a third of the evidence tokens and abstained on 90.00% of 30 unanswerable questions.

5 Conclusion

EdgeMem preserves original interaction turns and organizes them in a multi-anchor hypergraph for LLM-free construction and retrieval. Across LoCoMo and LongMemEval-S, it provides strong answer quality, retrieval, abstention, and efficiency while keeping answers traceable to source evidence.

  • Conclusion: EdgeMem stores original interaction turns in Time, Co-occurrence, and Episode Sub-Hypergraphs, then fuses their outputs into one evidence pack.Retrieval activates query-matched hyperedges, and a reader composes each answer from the resulting source evidence.
  • Conclusion: On LoCoMo, EdgeMem achieves the strongest answer quality under a shared prompt and abstains more reliably when the conversation does not support an answer.The conclusion reports this alongside the method’s source-grounded design.
  • Conclusion: On LongMemEval-S, the fixed configuration obtains the highest session-level Recall@3 and NDCG@3 in the reported suite.These are benchmark-level session retrieval measures.
  • Conclusion: Across both benchmarks, memory construction and retrieval use zero generative-LLM tokens, yielding efficient, source-grounded agent memory.Generative-LLM use is reserved for the final answer reader.

Ethical Considerations

The evaluation uses two publicly released research benchmarks and conducts no new data collection or human-subject study. Because EdgeMem preserves source turns verbatim, deployments should address consent, access, and deletion outside this evaluation.

  • Ethical Considerations: The study uses LoCoMo and LongMemEval-S, which contain synthetic or curated conversations, and conducts no new data collection or human-subject study.Datasets and reproduced systems are used under their released licenses.
  • Ethical Considerations: Real-world deployments should enforce consent-aligned retention, access-control, and deletion policies for verbatim source-turn storage.These deployment-level mechanisms are outside the scope of the evaluation.

A Prompts and Scoring Rules

The appendix documents the prompts, model settings, judge instructions, and adversarial scoring rules used to make evaluation reproducible and comparable. It covers EdgeMem’s reader, the shared answer prompt, strict and lenient judges, and category-5 handling.

  • A Prompts and Scoring Rules: All prompts are reproduced verbatim from released artifacts, with source identifiers and available SHA-256 checksums reported for each prompt.This appendix section documents the provenance of the evaluation prompts.
  • A.1 Reader Prompt: EdgeMem’s reader uses gpt-4o-mini-2024-07-18 at temperature 0 and instructs answers to rely only on supplied evidence, resolving dates against evidence-turn session dates.When evidence lacks the answer, the reader must emit an exact refusal string.
  • A.2 Shared Answer Prompt: The shared answer prompt uses retrieved texts and question metadata with gpt-4o-mini-2024-07-18 at temperature 0 and no refusal or length constraint.It is used for Table 1(b) and the style-control row of Table 4, with a 4,000-token maximum.
  • A.3 Strict Judge (SJ): The strict judge credits a response when it contains the reference answer or an equivalent answer with all required intermediate information.It returns only [[yes]] or [[no]] under a pinned gpt-4o-2024-08-06 snapshot.
  • A.4 Lenient Judge (LJ): The lenient judge accepts same-topic answers and same-period time references, and its table row orderings match those of the strict judge.On 20,013 paired verdicts, it upgrades 17.7% of strict-judge negatives and downgrades 0.1%.
  • A.4 Lenient Judge (LJ): LoCoMo’s lenient-judge rule excludes adversarial questions from its denominator, while the strict judge uses the released category-5 handling.The appendix separately specifies the lenient judge’s JSON label contract.
  • A.5 Adversarial Scoring Rule: Of 446 adversarial questions, 444 are correct only for the exact refusal string, while the two with reference answers are scored by the strict judge.No lexical metric is computed against an adversarial trap answer, and the rule was fixed before baseline runs.
Loading 2609.05553v1…