Memory-Autonomy Framework
Memory-autonomy framework is Su Yu / 苏煜’s compressed account of what makes an agent capable 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. Memory covers semantic knowledge, episodic memory, and procedural knowledge; autonomy covers perception, reasoning, decision, and action.
The framework lets the source compare otherwise different agent paradigms. Logical agents had explicit symbolic memory and inference but weak knowledge acquisition; neural game agents had stronger learned action in repeatable environments but limited explicit memory; Language Agent systems use language as a shared substrate for memory, reasoning, and tool-mediated action.
Key Claims
- Agent reliability depends on both what the system can remember and how independently it can act toward a goal.
- Persistent Agent Memory is only one part of memory; procedural knowledge and world-specific operational knowledge also matter.
- Agent Harness systems supply much of the autonomy layer by giving agents tools, environments, permissions, and feedback.
- Continual Learning becomes necessary when memory and autonomy need to improve from real work rather than from a fixed prompt.
Connections
- Language Agent — modern agent form interpreted through this framework.
- Semantic Parsing — route for translating human intent into machine-action representations.
- Persistent Agent Memory, AI Skills, and Agent Harness — wiki concepts that map onto the framework’s memory and autonomy sides.
- World Models and Specialized Intelligence — what agents must learn to become expert in specific environments.