MemGen: Weaving Generative Latent Memory for Self-Evolving Agents (2509.24704v1)

Abstract: Agent memory shapes how LLM-powered agents, akin to the human brain, progressively refine themselves through environment interactions. Existing paradigms remain constrained: parametric memory forcibly adjusts model parameters, and retrieval-based memory externalizes experience into structured databases, yet neither captures the fluid interweaving of reasoning and memory that underlies human cognition. To address this gap, we propose MemGen, a dynamic generative memory framework that equips agents with a human-esque cognitive faculty. It consists of a \textit{memory trigger}, which monitors the agent's reasoning state to decide explicit memory invocation, and a \textit{memory weaver}, which takes the agent's current state as stimulus to construct a latent token sequence as machine-native memory to enrich its reasoning. In this way, MemGen enables agents to recall and augment latent memory throughout reasoning, producing a tightly interwoven cycle of memory and cognition. Extensive experiments across eight benchmarks show that MemGen surpasses leading external memory systems such as ExpeL and AWM by up to $38.22\%$, exceeds GRPO by up to $13.44\%$, and exhibits strong cross-domain generalization ability. More importantly, we find that without explicit supervision, MemGen spontaneously evolves distinct human-like memory faculties, including planning memory, procedural memory, and working memory, suggesting an emergent trajectory toward more naturalistic forms of machine cognition.

• The paper introduces a generative memory model that interweaves reasoning with latent memory, outperforming benchmarks by up to 38.22%.
• The methodology uses a memory trigger via reinforcement learning and a memory weaver to generate context-rich latent token sequences.
• MemGen demonstrates robust cross-domain generalization and mitigates catastrophic forgetting, paving the way for human-like cognitive AI systems.

MemGen: Weaving Generative Latent Memory for Self-Evolving Agents

Introduction

The paper "MemGen: Weaving Generative Latent Memory for Self-Evolving Agents" proposes a novel memory framework to enhance LLM-powered agents. Current paradigms, such as parametric and retrieval-based memory, are restricted in imitating human-like cognition due to their inability to fluidly interweave reasoning and memory. MemGen introduces a highly dynamic generative memory model that mimics the human cognitive cycle, involving a "memory trigger" for detecting reasoning states that require memory recall and a "memory weaver" that constructs latent token sequences to enrich cognitive processing. Experimental results demonstrate MemGen's superiority, outperforming existing memory systems by significant margins and displaying emergent cognitive traits like planning and procedural memory.

Methodology

MemGen combines two main components: the memory trigger and the memory weaver. The memory trigger utilizes reinforcement learning to discern when memory should be invoked, relying on hidden state vectors generated during reasoning to estimate the likelihood of beneficial memory integration. Conversely, the memory weaver dynamically generates latent memory sequences when invoked, synthesizing external data with internal cognitive states for optimal problem-solving.

Figure 1: Overview of the proposed MemGen framework, illustrating interaction between memory trigger and weaver.

Experimental Results

MemGen significantly outperformed various benchmarks, such as the parametric and retrieval-based memory models. Across nine benchmarks, MemGen yielded up to a 38.22% improvement over the state-of-the-art memory systems. Additionally, it demonstrated robust cross-domain generalization and bolstered continual learning capabilities, reducing the catastrophic forgetting typically seen in parametric models. MemGen's ability to spontaneously develop human-like memory categories further showcases its potential in mimicking complex cognitive behaviors.

Analysis of Memory and Reasoning Interleaving

A core aspect of MemGen's design is the seamless interleaving of memory and reasoning, which effectively bypasses the static reliance seen in retrieval-based memory systems. As shown in Figure 2, t-SNE visualizations indicate clear clustering of similar memory types, suggesting that MemGen is proficient in organizing and utilizing memory in a meticulous, segmented manner akin to human cognition.

Figure 2: t-SNE visualization of latent memories generated by MemGen across different datasets.

Implications and Future Developments

The implication of MemGen's research is profound, indicating actionable pathways towards developing machines capable of fluid, context-aware reasoning reminiscent of human thought processes. As AI agents progressively evolve with these advanced memory paradigms, the potential applications span across complex decision-making environments, from autonomous vehicles to personalized healthcare systems. Future work could explore more granular integration of memory types and scaling mechanisms to expand the applicability of MemGen.

Conclusion

MemGen introduces a comprehensive and dynamic architecture for integrating memory into AI systems, overcoming the limitations of traditional paradigms by enabling seamless reasoning and recall cycles. Through extensive experimentation and analysis, MemGen proves to be a significant advancement in building AI agents that emulate human-like cognitive processes, exhibiting emergent memory faculties and adaptability across diverse tasks.