This paper addresses the limitations of modern language agents that retrieve external information, adapt to observations, and answer interdependent questions in long-term, multi-round interactions. Existing LLM systems rely on full context prompting, which appends all past rounds regardless of relevance, resulting in infinite memory growth, increased computational cost, and poor inference performance for input lengths outside the distribution. In response, this paper proposes MEM1, an end-to-end reinforcement learning framework that can perform long-term, multi-round tasks using constant memory. MEM1 updates a compressed shared internal state that supports memory integration and inference at each round, integrating new observations from the environment with previous memories while strategically removing irrelevant or redundant information. In addition, we propose a simple, yet effective, and scalable method to support learning in more realistic and constructive environments by composing existing datasets into arbitrarily complex task sequences. Experiments across three domains, including internal search QA, open-domain web QA, and multi-hop web shopping, demonstrate that MEM1-7B improves performance by 3.5x over Qwen2.5-14B-Instruct on a 16-objective multi-hop QA task while reducing memory usage by 3.7x, and generalizes well beyond the training period. Our results demonstrate the potential of inference-based memory integration as a scalable alternative to existing solutions for training long-term interacting agents that optimizes both efficiency and performance.
