Organizational Memory Enhances LLM Agent Business Processes

The research introduces the concept of an "organizational memory" to address a key limitation of LLM-based agents in enterprise settings: their lack of specific, internal company knowledge. While LLMs can automate business processes beyond traditional rule-based systems, they struggle with fragmented, human-oriented information like policies and standard operating procedures. Encoding this knowledge directly into prompts or individual agent setups is not scalable and leads to silos and inconsistencies. The proposed organizational memory acts as a shared, governed, and agent-consumable reference layer for evolving procedural knowledge. The paper derives essential requirements for such a system, presents an architecture for its curation and consumption, and validates its efficacy through a proof-of-concept in a procurement scenario. This approach aims to provide agents with consistent, up-to-date organizational context, enabling more reliable and scalable business process execution.