Paper 2606.06036
Memory is Reconstructed, Not Retrieved: Graph Memory for LLM Agents
- Published
- Jun 2026
- Research lab
- Independent
- Citations
- 2
- GitHub
- 233 stars
01 In brief
Summary
The paper introduces MRAgent, a framework for LLM agents that treats memory access as an active, multi-step reconstruction process over a structured memory graph, rather than a passive retrieve-then-reason pipeline.
MRAgent organizes memory as a Cue–Tag–Content graph, where associative tags link fine-grained cues to memory contents, enabling the LLM to iteratively explore and prune retrieval paths based on accumulated evidence.
This approach addresses limitations of existing memory systems, which rely on fixed top-k selection or predefined graph traversal and cannot adapt to intermediate findings.
Experiments on the LOCOMO and LongMemEval benchmarks show MRAgent outperforms strong baselines by up to 23% in LLM-Judge scores, while reducing token consumption and runtime.
A theoretical analysis proves that active retrieval policies are strictly more expressive than passive ones.
Ablation studies confirm that multi-step reasoning and associative tags are key contributors, and that episodic and semantic memory layers are complementary.
The work is motivated by cognitive neuroscience, which views memory retrieval as an active and associative reconstruction process.
02 From the paper
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.