Augmenting Language Agents with Parametric Memory

Large Language Models (LLMs) have demonstrated strong reasoning abilities, yet existing agent frameworks remain constrained by two limitations. First, they typically operate at the per-instance level, confining signals to individual problems and overlooking transferable patterns across tasks. Second, while some approaches attempt to incorporate global information through external memory, these are non-parametric in nature, and thus capture only shallow interactions across instances, failing to uncover deeper regularities. To overcome these limitations, we propose $\texttt{ParamAgent}$, a language agent framework that leverages a domain-adaptive parametric memory to internalize knowledge across samples into model parameters. In addition to capturing cross-sample regularities, $\texttt{ParamAgent}$ provides twofold flexibility: (i) the parametric module can supply different forms of knowledge depending on various domains, and (ii) the same module can be integrated with different base LLMs, making $\texttt{ParamAgent}$ broadly applicable. Moreover, $\texttt{ParamAgent}$ naturally promotes diversity of outputs by adjusting the sampling temperature of the parametric module. Experiments on programming, math reasoning, and multi-hop question answering benchmarks show that $\texttt{ParamAgent}$ consistently outperforms state-of-the-art baselines, surpassing the best baseline by up to $7.90\\%$, $9.41\\%$, and $24.30\\%$ respectively.