New agentic memory framework uses 118K tokens per query. LangMem burns through 3.26M.

2026-06-26 · Source: VentureBeat · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Robotics & Autonomous Systems · Depth: Advanced, medium

Summary

Researchers at the National University of Singapore developed MRAgent, a novel agentic memory framework designed to overcome the limitations of long-horizon reasoning in AI agents, where context windows quickly fill and retrieval pipelines yield noisy results. MRAgent departs from traditional "retrieve-then-reason" methods by dynamically developing memory through accumulating evidence, integrating multi-step memory reconstruction directly into the large language model's reasoning process. This framework treats memory as an interactive environment, exploring candidate retrieval paths across a structured "Cue-Tag-Content" memory graph and iteratively optimizing its search by inferring new constraints and pruning irrelevant branches. Benchmarked against systems like A-MEM and LangMem on LoCoMo and LongMemEval, MRAgent consistently outperformed baselines, reducing prompt token consumption to 118k per sample compared to LangMem's 3.26 million, and halving runtime to 586 seconds versus A-MEM's 1,122 seconds. Its efficiency stems from on-demand evaluation of tags and autonomous context assessment.

Key takeaway

For AI Engineers building agents for complex, long-horizon tasks, MRAgent offers a compelling alternative to traditional RAG. You should investigate integrating its active memory reconstruction to drastically cut token consumption and runtime costs. This approach allows your agents to dynamically refine memory access, avoiding context window overflow and improving reasoning accuracy. Consider implementing the automated distillation pipeline to efficiently populate the required Cue-Tag-Content memory graph.

Key insights

MRAgent dynamically reconstructs memory during reasoning, significantly reducing token use and improving long-horizon task performance.

Principles

• Active memory reconstruction outperforms passive retrieval.
• Dynamic memory development improves long-horizon reasoning.
• Iterative search optimization reduces context window noise.

Method

MRAgent uses an LLM to explore a "Cue-Tag-Content" memory graph, iteratively evaluating intermediate evidence to optimize search paths and prune irrelevant branches before retrieving full content.

In practice

• Implement automated LLM-driven memory graph population.
• Architect memory databases for associative navigation.
• Use MRAgent for complex, multi-turn agentic tasks.

Topics

• MRAgent
• Agentic AI
• Long-Horizon Reasoning
• LLM Memory Management
• Token Efficiency
• Graph Databases

Code references

• Ji-shuo/MRAgent

Best for: Research Scientist, AI Architect, NLP Engineer, AI Scientist, Machine Learning Engineer, AI Engineer

Related on AIssential

• Memory is Reconstructed, Not Retrieved: Graph Memory for LLM Agents — LLM agent memory should be actively reconstructed and dynamically adapted to reasoning context, not statically retrieved.
• A 0.12% parameter add-on gives AI agents the working memory RAG can't — Delta-mem offers an efficient, dynamically updated associative memory for AI agents, overcoming RAG and context window limitations.
• How xMemory cuts token costs and context bloat in AI agents — xMemory uses a hierarchical memory structure and uncertainty-gated retrieval to optimize long-term AI agent coherence and reduce token costs.
• rohitg00 / agentmemory — Agentmemory provides persistent, searchable memory for AI coding agents, enhancing efficiency and reducing redundant explanations.
• AMA: Adaptive Memory via Multi-Agent Collaboration — AMA uses multi-agent collaboration and hierarchical memory to provide adaptive, consistent long-term memory for LLM agents.
• AtomMem: Building Simple and Effective Memory System for LLM Agents via Atomic Facts — AtomMem enhances LLM agent memory by extracting atomic facts and organizing them hierarchically for stable, efficient long-term recall.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by VentureBeat.