General Agentic Memory Via Deep Research
B. Y. Yan, Chaofan Li, Hongjin Qian, Shuqi Lu, Zheng Liu
TL;DR
The paper tackles memory bottlenecks in AI agents caused by static, ahead-of-time memory that loses fine-grained information. It introduces General Agentic Memory (GAM) based on just-in-time compilation, with a lightweight Memorizer offline and a deep Researcher online that performs iterative planning, searching, and reflection over a page-store; optimization is framed with a min–max objective and end-to-end reinforcement learning. Formally, $c^* = \mathrm{argmin}_{\mathcal{C}^*} |c|$ s.t. $\mathcal{C}^* = \mathrm{argmax}_{\mathcal{C}} \mathrm{Agent}(task, context)$, and policy-gradient updates for $\theta_m$ and $\theta_r$ are derived as $\nabla_{\theta_m}$ and $\nabla_{\theta_r}$ in the paper. Empirically, GAM outperforms baselines on LoCoMo, HotpotQA, RULER, and NarrativeQA, exhibiting strong memory fidelity, effective long-context reasoning, and robustness to context length, with test-time scalability enabled by deeper reflection and larger page retrieval; the work also provides a public release to foster further research.
Abstract
Memory is critical for AI agents, yet the widely-adopted static memory, aiming to create readily available memory in advance, is inevitably subject to severe information loss. To address this limitation, we propose a novel framework called \textbf{general agentic memory (GAM)}. GAM follows the principle of "\textbf{just-in time (JIT) compilation}" where it focuses on creating optimized contexts for its client at runtime while keeping only simple but useful memory during the offline stage. To this end, GAM employs a duo-design with the following components. 1) \textbf{Memorizer}, which highlights key historical information using a lightweight memory, while maintaining complete historical information within a universal page-store. 2) \textbf{Researcher}, which retrieves and integrates useful information from the page-store for its online request guided by the pre-constructed memory. This design allows GAM to effectively leverage the agentic capabilities and test-time scalability of frontier large language models (LLMs), while also facilitating end-to-end performance optimization through reinforcement learning. In our experimental study, we demonstrate that GAM achieves substantial improvement on various memory-grounded task completion scenarios against existing memory systems.