Continual Learning
Continual learning is the agent-learning bottleneck emphasized by Su Yu / 苏煜 in 139. 【Agent的综述】和苏煜聊Agent技术史、OpenClaw Moment、边界的消弭和社会的辐射. The source notes that the term is used broadly, including avoiding task forgetting, personalization, self-improvement, post-training, and OpenClaw-like learning, but Su’s main question is what the agent is trying to learn.
His answer is that continual learning should help agents acquire broader World Models of concrete working environments: organization structure, workflows, software operation, interpersonal context, task history, and theory of mind. Without that learning, agents remain unreliable, slow, expensive, and hard to specialize.
为什么硅谷开始重新定义「AI 记忆」｜ S10E20 adds a caution from 康宏文 Henry: retraining or parameter-updating a model can personalize behavior, but it may not preserve precise personal memory. The episode distinguishes probabilistic preference learning from a Local-First Memory Layer that can recover exact older files, clips, claims, and timestamps.
Key Claims
- Continual learning is the shared bottleneck behind memory, self-learning, personalization, post-training, and expert-agent behavior.
- Learning should not merely accumulate raw history; it should build useful models of the small world where the agent operates.
- Persistent Agent Memory, AI Skills, and Agent Self-Evolution are product-level pieces of the same learning problem.
- Solving continual learning would make Specialized Intelligence and Universal Digital Agent more practical by reducing repeated instruction and failure.
- Continual learning and parameterized personalization should not be mistaken for complete personal memory when exact recall and source grounding matter.
Connections
- World Models — the source’s answer to what continual learning should learn.
- Specialized Intelligence — target capability enabled by continual learning.
- Persistent Agent Memory, AI Skills, and Agent Self-Evolution — related wiki concepts for retaining and reusing experience.
- Agent Post-Training and Agent RL — adjacent training approaches that can use agent traces and workflow feedback.
- Local-First Memory Layer and Data-to-Memory Transformation — S10E20’s external-memory contrast with parameterized learning.