Procedural Memory for Persistent LLM Agents: Know-How as the Missing Component

Key Points

• The research aims to address the lack of procedural memory in persistent LLM agents, which hampers their ability to apply learned procedures.
• Formalized procedural memory as consolidated action patterns.
• Developed an architecture for extraction, storage, and incremental refinement of these action patterns.
• Illustrated how this framework operates in context of repeated experiences.
• Proposed architecture facilitates the internalization of procedures for LLM agents.
• Demonstrated that agents could effectively handle repeated tasks without rediscovery.
• Indicated a path toward improved operational efficiency in LLM applications.

Abstract