Agent context files like AGENTS.md and CLAUDE.md have become the standard way to tell Claude Code, Cursor or Copilot how a project works. Almost nobody keeps them current, because writing down what you already know is the first task to get dropped. Meanwhile, PiecesOS has been quietly recording the decisions, the bugs and the dead ends the whole time. This CLI moves one into the other.
A run searches your memory over MCP with two strategies per category. Vector search understands paraphrase but cannot see what was indexed minutes ago; full-text search matches strictly and immediately but misses anything worded differently. Running both covers the gap. Matches are filtered to the current repository, redacted, and composed into markdown between two managed markers, so whatever you wrote by hand survives every run.
Nothing leaves your machine: every call goes to localhost, with no account and no telemetry. And nothing is written without you reading it first. Each run prints a diff and waits for approval, and there is deliberately no flag to skip that step.
Under the hood it reads the summary chain Pieces already writes (about 70 ms) instead of the ask_memory endpoint (about 3.3 s of raw screen OCR), so runs stay fast and the output stays readable markdown.