Memento 3: Model-Based Recursive Self-Improvement through Reflective Rulebooks
- 类型：arxiv
- 标识：2610.11794
- 链接：https://arxiv.org/abs/2610.11794
- 主分类：agent
- 形态：method
- TLDR：Learning to act in unfamiliar environments requires agents to infer how the world works and revise that understanding as new evidence arrives. Yet limited observations can support multiple world models that explain past interactions but predict different outcomes in unseen states. We introduce Memento 3, building on the Memento series to enable frozen LLM agents to continually learn explicit world models through external memory. The agent maintains a natural-language rulebook as persistent semantic memory, recording revisable hypotheses about environment dynamics while leaving unknown aspects
- 待LLM分类：否
- 标题中文：Memento 3：通过反思式规则手册实现基于模型的递归自我改进
- TLDR中文：在未知环境中学会行动需要智能体推断世界运作方式，并在新证据出现时修正理解。然而，有限的观察可能支持多个世界模型，它们都能解释过去的交互，但对未见状态有不同的预测。我们提出 Memento 3，在 Memento 系列的基础上，使冻结的 LLM 智能体能够通过外部记忆持续学习显式世界模型。智能体维护一个自然语言规则手册作为持久化语义记忆，记录可修正的环境动态假设，同时保留未知部分
- 来源文件：
- /inbox/tom/_candidates/2026-10-09-agent-rag-longcontext-candidates.json