MemPO: Self-Memory Policy Optimization for Long-Horizon Agents

Researchers have introduced MemPO, a self-memory policy optimization algorithm designed to help long-horizon artificial intelligence agents manage their own memory during tasks, according to a paper posted on arXiv [1]. The algorithm, detailed by Ruoran Li and colleagues, addresses a core problem for agents that interact with environments over extended periods: the accumulation of context degrades performance and stability [1]. Existing approaches often rely on an external memory module that the agent queries, but this prevents the model from proactively curating its memory in alignment with its objectives [2]. MemPO instead enables the policy model to autonomously summarize and manage its memory as it works [1]. It does so by improving the credit assignment mechanism based on memory effectiveness, allowing the agent to selectively retain crucial information [2]. In experiments, MemPO delivered an absolute F1 score gain of 25.98 over the base model and 7.1 over the previous state-of-the-art baseline, while cutting token usage by 67.58 percent and 73.12 percent, respectively [1]. The code has been released on GitHub [2]. The paper appeared on arXiv, an open-access repository for electronic preprints that, as of late 2024, receives about 24,000 submissions per month across fields including computer science and artificial intelligence [6]. Submissions to arXiv are moderated but not peer-reviewed [6]. The MemPO manuscript was first submitted on February 28, 2026, and revised most recently on June 15, 2026 [1].