Unified Agentic Memory Across Harnesses Using Hooks

A unified memory layer can be created for different AI coding agents to avoid vendor lock-in and provide persistent knowledge. The proposed architecture uses standardized 'hooks' to passively log all session events from harnesses like Claude Code, Codex, and Cursor into a Neo4j graph database. This deterministic logging captures a complete audit trail without consuming the agent's attention or context. An offline 'dream phase' process then uses an LLM to periodically analyze these event logs, distilling them into durable, structured memories that can be injected back into new sessions.