Memory is Reconstructed, Not Retrieved: Graph Memory for LLM Agents

2026-06-04

AI Summary

Existing LLM agents are constrained by static “retrieve-and-reason” memory paradigms when handling long interaction histories, limiting their ability to dynamically adjust memory access based on intermediate evidence. This work proposes MRAgent, a framework that models memory as a Cue-Tag-Content graph structure and introduces an LLM-guided active reconstruction mechanism, deeply integrating memory access with the reasoning process. By enabling iterative path exploration and pruning grounded in accumulated evidence, MRAgent treats memory as a dynamic, reconstructive process rather than a fixed repository—departing fundamentally from conventional static retrieval. Evaluated on the LoCoMo and LongMemEval benchmarks, the method achieves up to a 23% performance gain over strong baselines while significantly reducing token consumption and runtime.

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.

Problem

• LLM agents
• long-horizon reasoning
• memory retrieval
• dynamic memory access
• reasoning over interaction histories

Innovation

• associative memory graph
• active reconstruction
• dynamic memory access
• Cue-Tag-Content graph
• long-horizon reasoning