Self-evolution
Corrections become learnings that load every session, and a monthly review, run with you, prunes the brain and turns recurring signals into prompt and skill edits.
Your agent gets better at working with you in two ways. Every correction becomes a learning it reads at the start of each session. And once a month it sits down with you to clear out what has gone stale and to edit its own instructions.
Feedback is durable
Every correction you give mid-conversation ("don't do that", "actually, that's wrong") lands in knowledge/raw/user/feedback.md, append-only. The next compile synthesizes all feedback into a Learnings block in knowledge/memory/index.md, which loads at the start of every future session. The agent sees its own past misses while it works. To log a correction yourself, see Capture.
Self-review, monthly
Stale, wrong and crowded data costs as much as missing data: an outdated fact gets acted on, and it buries what's live. self-review is the monthly upkeep for both, run with you, in two passes. Nobody remembers a monthly chore, so sync offers it when it's due, first and on its own: run it now, tomorrow, or skip this month.
- The brain pass comes due every 30 days. A new home is first offered one once its oldest session is a month old.
- The rules pass comes due when new feedback has piled up and the last pass is over 14 days old, or after 30 days regardless.
A bare self-review runs both, brain first. self-review brain or self-review rules runs one.
The brain pass
A quick inventory finds what has gone stale, and the agent asks you about each item with its evidence and a recommendation, four questions at a time, applying every batch before the next. After 24 questions you choose to continue or stop; whatever's left comes first next month.
| What it finds | What you can answer |
|---|---|
| Tasks untouched for 30 days (dormant when nobody has mentioned them in 60) and active tasks over 60 days old | Keep this month · Park · Cancel · Already done |
| Projects with no task touched and no mention in 60 days | Keep · Pause · Archive |
| Memory lines whose tasks closed or went quiet, pending items, and key facts | Drop · Keep · Rewrite |
| Recently archived decisions that no permanent article holds yet | Promote into an article · Leave archived |
| Articles that current state has outrun | Rewrite · Still accurate |
| Pinned reports whose file is gone | Fix the link · Unpin |
| The roadmap, a project's dashboard, or a pinned report unchanged for 60 days | Refresh · Still accurate · Unpin or Promote into the brain (pins only) |
| Screenshots and captures older than a month, grouped by month | Keep · Delete |
A Keep stays quiet for 90 days while the item is unchanged, so the same question doesn't come back next month. Every answer is applied in the session, and every deleted or rewritten line is quoted in the report.
The rules pass
The rules pass works on the agent's instructions and the feedback loop. It casts a wide net: the Learnings synthesis, the full feedback log, recent sessions and task threads, git history, concept articles, in-flight plans, and installed skills. It greps for both correction and confirmation phrases; wins validate non-obvious choices.
It prunes before it adds. Everything loaded at session start is paid for on every turn, so a line earns its place only if deleting it would change behavior. Candidates are rules already graduated into the manual, rules a later decision superseded, references to things that no longer exist, prose a hook or test now enforces, and generic advice the current model already follows. That last kind is retired on trial: removed, then restored if it's broken again. Nothing is deleted without your approval.
Then each recurring theme is classified as missing, buried, present but violated, or present and working. The point is catching rules that landed but did not stick. Themes rank by severity × instances × cycles; anything with fewer than two independent signals is dropped. If nothing clears the bar, the agent says so and stops.
Proposals come in five tracks:
| Track | What | Applied |
|---|---|---|
| A | Prompt or skill edits. Identity → SOUL.md, procedural → AGENTS.md, skill-specific → that skill | In session, you pick which to accept |
| B | Code-change plans written to reports/plans/ | Never auto-applied; you implement them in a separate session |
| C | Install or create a skill, only for a recurring multi-step procedure | Explicit approval required |
| D | Promote a generic fix to the plugin, after a check that strips anything personal | Only after you approve the exact change |
| E | Compare this home with the plugin's templates; generic rules go upstream, personal ones are recorded and skipped | Every cycle, as a batch |
Aging proposals (over 14 days, no follow-through) get re-surfaced, downgraded, or closed. A quality gate runs before anything lands: every target file actually read, specific evidence with timestamps and quotes, and for any pre-existing rule, violations counted after it was introduced.
Capture is automatic, compile is on demand, and review is offered monthly and run with you. The brain covers the first two.
Sync
One pass through every maintenance step, in dependency order, ending in a status block and, when something needs you, a few closing questions.
Tax and books
Know what tax you owe at every moment, for your companies and your personal return. Every obligation becomes a dated task, receipts land in one ledger, the month closes for your accountant, and the estimate is planned before the penalty, not after.