As LLM agents move from short isolated interactions to sustained deployment, their effectiveness increasingly depends on memory that can retain important information, adapt to evolving users and environments, and make use of what has been learned from past interactions, but realistic workloads make it difficult to balance reliable fact preservation with the reuse of accumulated experience. To address this challenge, we propose Mixture of Memory (MoMem), a hybrid architecture that separates memory for facts from memory for reusable experience into textual and parametric pathways. During inference, a router decides which textual records to inject and whether to activate parametric memory using logic-matched trajectory evidence; after each interaction, a decoupler separates factual signals from reusable reasoning experience and updates the two pathways independently. Experiments show that, in the main mixed-workload setting, MoMem performs strongly on both sides of this balance, reaching 51.4 on LongMemEval, 11.7 points above the strongest baseline, and 86.4 on MATH-500, 3.2 points above the backbone model. Ablation studies show that these gains come from routing and decoupling rather than hybrid storage alone. Code is available at https://github.com/ShuoSun-CN/MoMem_a