Large language models have a memory problem. When you ask them to handle long-term, complex tasks, their context windows quickly fill up with useless noise, driving up computing costs. Now, researchers at the National University of Singapore (NUS) have built a workaround. They introduced MRAgent, a new framework that drops token consumption to just 118,000 tokens per query. That is a sharp drop from the industry standard LangMem framework, which burns through 3.26 million tokens for similar tasks.
This 96 percent reduction changes the math for running autonomous systems. The framework works by abandoning the old static method of retrieving data and then reasoning over it. Instead, MRAgent uses an active reconstruction process inspired by cognitive neuroscience. The AI builds a multi-granular associative memory graph. It evaluates semantic cues and tags on the fly rather than pulling raw text blindly.
As the agent gathers evidence, it actively prunes irrelevant memory branches. It stops the unconstrained expansion of the context window. It only holds onto the data it strictly needs to solve the immediate problem. This keeps runtime computing costs incredibly low while allowing the agent to function over longer periods. This is a clear upgrade over early retrieval-augmented generation designs that choked on long interactions, according to the NUS research team.
Why the GitHub Release Lowers the Barrier for Enterprise Startups
The true value of this 118K token efficiency is access. The researchers made the MRAgent framework and its code publicly available. This open-source distribution allows smaller enterprise companies to deploy long-horizon AI agents without paying the prohibitive computing fees previously demanded by the industry.
By removing the cost barrier, this open-source GitHub release lets startup ecosystems compete directly with larger infrastructure players in 2026. Startups can now experiment with complex autonomous tasks without burning through their operational budgets on excessive token limits.