OKF Agent Memory introduces a Git-native, domain-neutral persistent memory solution for AI agents based on the Open Knowledge Format (OKF) v0.2. To solve the issues of context window bloat caused by "Prompt Monoliths" and the semantic retrieval failures of RAG, the project implements a Dual-Memory Agent Architecture (DMAA). This architecture bifurcates memory into a "Push Layer" (Normative Working Memory), which utilizes an ultra-compact Agent Action Grammar (AAG) to enforce behavioral invariants via a 100–150 token codex, and a "Pull Layer" (Semantic Domain Memory), which provides on-demand access to a knowledge bundle without increasing baseline system prompt tokens. Developed in Go, the system achieves high performance with in-memory BM25 retrieval times under 300µs and graph validation in approximately 4ms. Unlike traditional vector database approaches (e.g., Mem0 or Letta), OKF operates with zero API costs and minimal memory overhead (<15 MB RSS), utilizing local lexical indexing rather than expensive embedding services. The implementation is entirely Git-native, storing all knowledge as plain text for version control, and includes a Model Context Protocol (MCP) server to integrate with platforms like Claude Code and Cursor.