what the project already learned, across context windows

Compaction protects the model's context window. It also throws away what was already tried, which is why an agent can cheerfully re-attempt a fix that failed an hour ago. DGC keeps the durable half separately: short, typed notes about this project, written as the work happens.

They are per project, so a run spread over days accumulates one trace, and they are written by the harness from tool results it already sees — no model cooperation required, so the same behaviour holds on any local endpoint.

• failure — a command or an edit that failed, with its exit code and the tail of its output.
• outcome — a file that was written or patched; a test command that passed.
• requirement / decision — from your standing goal and from a compaction summary.

Every note is bounded, redacted with the same secrets a saved session is, and carries the file, the tool and the date it came from, so a stale note can be judged rather than trusted.

• /notes — the most recent notes. /notes <query> searches the whole trace: an error string, a test name, a file path. dgc notes [QUERY] does the same from a shell, and DGC: Show Context Notes runs it from the editor.
• The agent has a notes tool with the same reach, so it can check whether something already failed before trying it again. It is read-only and allowed in plan mode.
• After a compaction, a bounded digest of requirements, decisions and recent failures is carried into the fresh context — the point at which this history would otherwise be lost.
• While editing, a file that has failed before carries that history into the result the agent reads, once per file per turn, so the next attempt is not the same one again.

/notes off stops recording everywhere, and notes: false in the config does the same. The store lives beside the session transcripts (~/.dgc/sessions/<project>/notes.sqlite), holds the most recent notes_max_rows entries, and is never sent anywhere.

This is not the /recall archive. That is your raw scrollback, kept for you to read and never put back into the model's context. Notes are a small curated projection that deliberately can be — which is why they are bounded and redacted.