HiMA-Ecom tests how shopping AI agents handle memory over long sessions

Context grows by accumulation

As arxiv.org reported, the paper starts from a simple observation. An agent system does not read one huge document. Instead, every step rebuilds a prompt from role instructions, tool descriptions, recalled memories, dialogue history and earlier tool outputs, separately for each agent.

So the context, the text a model has to read before answering, keeps growing as the task goes on. The paper studies how to train, test and manage that growing pile in assistants made of a manager agent and sub-agents.

What the benchmark contains

HiMA-Ecom is a benchmark for e-commerce assistants with 22.8K instances. Of these, 17.7K carry memory entries in the context, so the agents must use what was stored earlier.

It combines two kinds of training material. One is supervised fine-tuning data for individual agents, meaning examples of correct behavior. The other is system-level supervision for training the whole team end to end with reinforcement learning, where agents learn from rewards.

How HiMA-R1 keeps context in check

The authors build a system called HiMA-R1 on top of the benchmark. It uses a training method named VR-GRPO. It samples along an initial trajectory, so the sampling cost grows additively with trajectory length, and it updates the agents with the largest reward variance.

It also adds a reward for shorter trajectories, plus a memory evolution mechanism. That mechanism reuses the existing training rewards as free signals for deciding what stays in context.

The reported results

According to the abstract, agents built on 3B and 7B open-source models match DeepSeek-R1 and surpass DeepSeek-V3 by 6.0 points on average. Reward-driven memory also reduces the update steps needed for convergence by roughly 55%.

These are the authors' own claims from the abstract. This is a revised version of an existing paper, and we have not checked its methods or whether the data is public.