KHMS – a file-based long-term memory an LLM agent installs into itself (github.com)

KHMS (Know-How Management System) introduces a novel long-term memory solution for LLM (Large Language Model) agents, employing a file-based system utilizing markdown files organized within a git repository. Each knowledge entry is represented as a unique, immutable "card" containing metadata in YAML format, detailing the type of knowledge, its evidence strength, and the sources of derivation. This immutability ensures corrections are made through newly created cards instead of edits or deletions, which retains historical context—a feature crucial for agent learning and performance. The system's retrieval mechanism enhances efficiency by automatically inserting relevant cards into the agent's session and employs a rigorous review pipeline to validate new entries. This development is significant for the AI/ML community as it addresses critical flaws in traditional knowledge storage approaches, such as context loss and incomplete recall associated with short-session contexts. KHMS's structured methodology promotes better traceability and reliability of knowledge claims, thus improving the agent's decision-making processes. By combining elements from existing standards like Google Cloud's Open Knowledge Format and leveraging concepts from prominent research papers, KHMS lays the groundwork for a persistent, interconnected memory system that evolves over time, maximizing the potential of LLM agents in complex environments.