Building your own Claude Code? The one thing the leak didn't include.

The Claude Code source leaked on March 31. 512,000 lines of TypeScript. Within 48 hours there were Python rewrites, Rust ports, and model-agnostic forks running on OpenAI, Gemini, and DeepSeek. One thing none of the forks include: persistent agent memory across sessions. What the forks are missing Claude Code's session memory — CLAUDE.md, project context — is static. It does not adapt. It does not track drift. It does not remember that the agent decided last Tuesday that your auth module was fragile, or that it changed its approach to error handling after a bad refactor. Anthropic solved this internally with KAIROS, their proprietary agent identity system. The forks are removing it. They are right to remove it. But nothing is replacing it. The result: every session starts cold. The agent has no continuity beyond what you put in CLAUDE.md by hand. What you actually need Three things: Snapshot — freeze the agent's current understanding at a point in time, hash-verified Drift detection —