AI agents need dynamic memory—here's how MRAgent cuts costs by 97%

AI agents struggle with long-horizon reasoning because their context windows quickly overflow with noise, forcing developers to choose between expensive memory bloat or shallow responses. Researchers at the National University of Singapore have introduced MRAgent, a new framework that reimagines how agents build memory during reasoning rather than before it.

MRAgent abandons the traditional "retrieve-then-reason" pipeline, instead enabling agents to reconstruct their working memory dynamically based on accumulating evidence. Unlike static retrieval systems, this approach allows the framework to revise its search strategy mid-reasoning, discard irrelevant data early, and focus on high-value information. The result is a system that processes complex queries with just 118,000 tokens per request—3.26 million tokens less than some alternatives.

Why static retrieval fails in long-horizon tasks

Most agentic systems rely on passive retrieval pipelines that pull documents via vector search or graph traversal before passing them to an LLM for reasoning. This method introduces three critical flaws that degrade performance:

• No mid-reasoning adaptation: If an agent retrieves a document missing a key detail like a date or person’s name, it cannot issue a follow-up query to fill the gap.
• Irrelevant noise overload: Fixed similarity scores and predefined graph paths return superficial matches that flood the LLM’s context with extraneous information, sabotaging reasoning quality.
• Rigid structural dependencies: Systems depend on pre-optimized structures like top-k results and static relevance functions, which struggle to scale with unpredictable, real-world user interactions.

The researchers propose shifting to an "active and associative reconstruction process," drawing inspiration from cognitive neuroscience. Instead of treating memory as a static database, this approach treats it as an iterative discovery process where agents follow metadata stepping stones—small, contextual clues like names or actions—to piece together accurate narratives.

How MRAgent reconstructs memory in real time

MRAgent (Memory Reasoning Architecture for LLM Agents) frames memory not as a database but as an interactive environment. When processing a query, the system leverages the LLM’s reasoning capabilities to explore multiple retrieval paths across a structured memory graph. At each step, it evaluates intermediate evidence to refine its search, infer new constraints, prioritize promising paths, and prune irrelevant branches.

To achieve this efficiently, the framework organizes its data using a "Cue-Tag-Content" mechanism, a three-layer associative graph that separates retrieval into distinct stages:

• Cues: Fine-grained keywords extracted from user interactions, such as entities or contextual attributes.
• Tags: Semantic bridges that summarize relational connections between Cues and Content, enabling quick relevance judgments without heavy computation.
• Content: The stored memory units, divided into multi-granular layers like episodic memory (concrete events) and semantic memory (stable facts and preferences).

This structure enables a two-stage retrieval process that minimizes token waste. First, the LLM navigates from Cues to candidate Tags, using these summaries to evaluate relevance before accessing detailed content. Only the most promising paths proceed, drastically reducing unnecessary data ingestion.

Consider a user asking, "How did Nate use the prize money after winning his third video game tournament?" MRAgent:

1. Extracts initial cues such as "Nate," "video game tournament," and "win."
2. Maps these cues to memory graph Tags like "Tournament Victory" and "Tournament Participation."
3. Discards irrelevant tags (e.g., participation) and focuses on "Tournament Victory."
4. Retrieves episodic memories linked to Nate’s tournament wins, filtering down to the most relevant episode.
5. Updates its cues with new details (e.g., "tournament earnings") and repeats the process until it assembles a coherent answer—such as "Nate saved the money."

Benchmark performance and competitive landscape

MRAgent competes with frameworks like A-MEM, a graph-based agentic memory system, and MemoryOS, a hierarchical memory approach. Other persistent memory solutions include LangMem and Mem0, each targeting different trade-offs between accuracy, token efficiency, and runtime speed.

In industry benchmarks, MRAgent demonstrates a 97% reduction in token consumption compared to frameworks using static retrieval pipelines. This efficiency translates to faster response times and lower operational costs—critical advantages for applications requiring deep, long-horizon reasoning.

The framework’s active reconstruction mechanism also improves accuracy by avoiding the noise accumulation that plagues passive systems. By letting the LLM guide memory exploration, MRAgent aligns retrieval with reasoning goals, producing more reliable and contextually appropriate responses.

The future of agentic memory

As AI agents tackle increasingly complex, multi-step tasks, the demand for smarter, more efficient memory systems will grow. MRAgent represents a shift from rigid, pre-optimized retrieval to dynamic, reasoning-integrated memory reconstruction—an approach that could redefine how agents handle long-term context.

While challenges remain—such as scaling to massive knowledge graphs and handling ambiguous user queries—the framework’s early results suggest a promising direction. Developers building next-generation AI assistants, research agents, and automation tools may soon find that active memory isn’t just optional—it’s essential.