Follow AGENTS.md for this repo.
Before substantial planning or code changes, scan memory/impressions.jsonl for familiar cues. If familiarity is weak, read memory/summary.md. If familiarity is strong, read only the 1-3 linked detailed records needed for this task.
Do not load the whole memory tree unless I explicitly ask.
After the task, write back only durable decisions, architecture intent, stable preferences, unresolved follow-ups, milestone summaries, or recall feedback that should improve future cue quality.
Never store secrets, full transcripts, customer/user PII, low-signal chatter, raw logs, or disposable exploration noise.