Most agent setups handle memory badly. They either write everything to long-term memory until it fills with noise and contradictions, or they forget across sessions and you start from scratch every time. I have been building an open-source agent architecture (Apache-2.0) where memory is the part it tries hardest to get right, and where the same setup runs on Claude Code, Codex, or Gemini CLI instead of being locked to one tool. The core idea is that an agent should be a repo, not a prompt. The...