English

MemGen: Weaving Generative Latent Memory for Self-Evolving Agents

Computation and Language 2025-10-14 v2

Abstract

Agent memory shapes how Large Language Model (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%38.22\%, exceeds GRPO by up to 13.44%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.

Keywords

Cite

@article{arxiv.2509.24704,
  title  = {MemGen: Weaving Generative Latent Memory for Self-Evolving Agents},
  author = {Guibin Zhang and Muxin Fu and Shuicheng Yan},
  journal= {arXiv preprint arXiv:2509.24704},
  year   = {2025}
}
R2 v1 2026-07-01T06:04:24.783Z