MRAgent uses graph memory to improve LLM long-horizon reasoning
Researchers propose MRAgent, a framework combining an associative memory graph with an active reconstruction mechanism that improves LLM agent performance on long interaction histories by up to 23% over strong baselines while reducing token and runtime costs.
Score breakdown
MRAgent demonstrates that replacing static retrieval pipelines with evidence-guided, iterative graph traversal yields large accuracy gains on established long-horizon memory benchmarks while simultaneously cutting computational cost.
- 01Authors: Shuo Ji, Yibo Li, and Bryan Hooi
- 02Current memory-augmented agents rely on a static retrieve-then-reason paradigm that cannot adapt mid-inference
- 03MRAgent represents memory as a Cue-Tag-Content graph with associative tags as semantic bridges
Shuo Ji, Yibo Li, and Bryan Hooi identify a fundamental limitation in current memory-augmented LLM agents: their reliance on a static retrieve-then-reason paradigm that fixes memory access before reasoning begins, preventing the agent from dynamically adjusting what it retrieves as intermediate evidence emerges during inference. To address this, they introduce MRAgent, a framework that treats memory not as a static store to be queried once, but as a structure to be actively reconstructed during reasoning.
In this representation, associative tags serve as semantic bridges that connect fine-grained cues — the entry points for retrieval — to the underlying memory contents.
MRAgent encodes memory as a Cue-Tag-Content graph. In this representation, associative tags serve as semantic bridges that connect fine-grained cues — the entry points for retrieval — to the underlying memory contents. The framework's active reconstruction mechanism integrates LLM reasoning directly into the traversal of this graph, enabling the agent to iteratively explore promising retrieval paths and prune unproductive ones based on accumulated evidence. This design keeps retrieval contextually aligned with the evolving reasoning state while preventing the combinatorial explosion that would result from unconstrained graph expansion.
Experiments on the LoCoMo benchmark and the LongMemEval benchmark show improvements of up to 23% over strong baselines. Notably, these accuracy gains come alongside substantial reductions in both token usage and runtime cost, suggesting that the active, evidence-guided pruning strategy is more efficient than exhaustive retrieval approaches.
Key facts
- 01Authors: Shuo Ji, Yibo Li, and Bryan Hooi
- 02Current memory-augmented agents rely on a static retrieve-then-reason paradigm that cannot adapt mid-inference
- 03MRAgent represents memory as a Cue-Tag-Content graph with associative tags as semantic bridges
- 04An active reconstruction mechanism integrates LLM reasoning directly into memory access
- 05The agent iteratively explores and prunes retrieval paths based on accumulated evidence
- 06Evaluated on the LoCoMo and LongMemEval benchmarks
- 07Achieves up to 23% improvement over strong baselines while reducing token and runtime costs
Topics
Summary and scoring are generated automatically from the original article. We always link back to the publisher and never republish images or paywalled content. Last processed Jun 9, 2026 · 17:05 UTC. How this works →