MRAgent: 记忆是重建而非检索LLM Agent 的图记忆架架构

Source: https://arxiv.org/abs/2606.06036

Content fetched: 2026-08-02T12:19:11+08:00

Grounding: opencli arxiv paper

一句话

MRAgent 用 Cue-Tag-Content 关联图和主动重建机制替代静态检索，在两个基准上提升 23%

关键信息

arXiv ID: 2606.06036
English title: Memory is Reconstructed, Not Retrieved: Graph Memory for LLM Agents
Authors: Shuo Ji, Yibo Li, Bryan Hooi
Submitted/Published: 2026-06-04
Updated: 2026-06-04
Subjects: cs.AI, cs.IR
PDF: https://arxiv.org/pdf/2606.06036v1
Tags: agent-memory, graph-memory, retrieval, reasoning, llm-agent
Quality score: 4

中文摘要

MRAgent 挑战了记忆增强 LLM Agent 中的静态先检索后推理范式它将记忆表示为 Cue-Tag-Content 关联图，标签作为语义桥梁连接细粒度线索与记忆内容主动重建机制将 LLM 推理直接集成到记忆访问中，允许 Agent 迭代探索和修剪检索路径在 LoCoMo 和 LongMemEval 基准上超越强基线高达 23%，同时大幅降低了 token 和运行时间成本

English Summary

MRAgent challenges the static retrieve-then-reason paradigm in memory-augmented LLM agents. It represents memory as a Cue-Tag-Content associative graph where tags serve as semantic bridges connecting fine-grained cues to memory contents. An active reconstruction mechanism integrates LLM reasoning directly into memory access, allowing the agent to iteratively explore and prune retrieval paths based on accumulated evidence. Experiments on LoCoMo and LongMemEval benchmarks show up to 23% improvement over strong baselines while substantially reducing token and runtime cost.

Why it matters

MRAgent 用 Cue-Tag-Content 关联图和主动重建机制替代静态检索，在两个基准上提升 23%

arXiv Abstract

Despite recent progress, LLM agents still struggle with reasoning over long interaction histories. While current memory-augmented agents rely on a static retrieve-then-reason paradigm, this rigid pipeline design prevents them from dynamically adapting memory access to intermediate evidence discovered during inference. To bridge this gap, we propose MRAgent, a framework that combines an associative memory graph with an active reconstruction mechanism. We represent memory as a Cue-Tag-Content graph, where associative tags serve as semantic bridges connecting fine-grained cues to memory contents. Operating on this structure, our active reconstruction mechanism integrates LLM reasoning directly into memory access, allowing the agent to iteratively explore and prune retrieval paths based on accumulated evidence. This ensures that memory retrieval is dynamically adapted to the reasoning context while avoiding combinatorial explosion caused by unconstrained expansion. Experiments on the LoCoMo benchmark and LongMemEval benchmark demonstrate significant improvements over strong baselines (up to 23%), while substantially reducing token and runtime cost, highlighting the effectiveness of active and associative reconstruction for long-horizon memory reasoning.

Source Metadata

{
  "id": "2606.06036",
  "title": "Memory is Reconstructed, Not Retrieved: Graph Memory for LLM Agents",
  "authors": "Shuo Ji, Yibo Li, Bryan Hooi",
  "abstract": "Despite recent progress, LLM agents still struggle with reasoning over long interaction histories. While current memory-augmented agents rely on a static retrieve-then-reason paradigm, this rigid pipeline design prevents them from dynamically adapting memory access to intermediate evidence discovered during inference. To bridge this gap, we propose MRAgent, a framework that combines an associative memory graph with an active reconstruction mechanism. We represent memory as a Cue-Tag-Content graph, where associative tags serve as semantic bridges connecting fine-grained cues to memory contents. Operating on this structure, our active reconstruction mechanism integrates LLM reasoning directly into memory access, allowing the agent to iteratively explore and prune retrieval paths based on accumulated evidence. This ensures that memory retrieval is dynamically adapted to the reasoning context while avoiding combinatorial explosion caused by unconstrained expansion. Experiments on the LoCoMo benchmark and LongMemEval benchmark demonstrate significant improvements over strong baselines (up to 23%), while substantially reducing token and runtime cost, highlighting the effectiveness of active and associative reconstruction for long-horizon memory reasoning.",
  "published": "2026-06-04",
  "updated": "2026-06-04",
  "primary_category": "cs.AI",
  "categories": "cs.AI, cs.IR",
  "comment": "Accepted at ICML 2026",
  "pdf": "https://arxiv.org/pdf/2606.06036v1",
  "url": "https://arxiv.org/abs/2606.06036"
}