Skip to main content

​
Wolffish Sleeps On It

Memory records what happened. The feedback loop records what worked at the tool level. Reflection closes the last gap: every night, Wolffish reviews its own finished conversations the way a careful colleague reviews their day — what was attempted, how it actually ended, what worked well enough to repeat, what failed and why, and what you showed it about how you like things done. The lessons distill into a playbook that rides into every future conversation. The strongest signal in that review is yours: a 0–10 turn score, given with one click in the app or by texting a bare number from your phone.

​
Location

Like everything else in the workspace, these are plain markdown — open the playbook in any editor to see exactly what Wolffish has taught itself. reviewed.json entries are pruned after 14 days.

​
The Nightly Review

The nightly job looks back over the last 7 days and reviews every conversation — automations and procedure runs included — that has been quiet for at least the configured window (default 12 hours). A conversation that gets continued after its review is simply re-reviewed the next night. Each conversation gets one review call on your configured Brain (reasoning off, billed like the other utility side-calls, never against the conversation’s own meter). The result is an honest block appended to the day file:
The score: line records two things: the average of your 0–10 votes on the conversation’s turns (none when you didn’t vote), and the model’s own honest self-score — kept internal to reflection, never shown in chat.

​
The Playbook

The same nightly run folds new review blocks into playbook.md — a living document rewritten whole, never an append-only log. It has exactly five sections:
Every entry carries its provenance and date — user-said (you stated it), user-scored (your votes proved it), or inferred (the review concluded it). The merge rules keep it trustworthy:
  • Newest evidence wins a contradiction — a re-learned preference replaces the stale one instead of sitting beside it.
  • Inferred entries decay after 30 days unless reinforced; what you said yourself doesn’t fade.
  • Hard ceiling of 6,000 characters — the playbook must stay a briefing, not an archive.
  • Every rewrite keeps the previous version as playbook.md.bak, so a bad night is one copy away from undone.
The playbook rides in every system prompt, framed as tendencies rather than laws — your live instruction in the conversation always outranks it. Once the first playbook exists, it takes over the prompt slot previously held by the learned-preferences digest, which remains as the bootstrap fallback.

​
Your 0–10 Becomes Its Ground Truth

Scoring is the fastest way to teach the reflection loop what good looks like — and it works from every surface: The plumbing stays careful on the channels:
  • Digits typed on an Arabic keyboard (٠–٩ and ۰–۹) count the same as Western digits.
  • A “3” answering a numbered question card or /resume picker still means option 3 — the vote check runs only after pending pickers are ruled out.
  • A number opening a fresh conversation is just a message.
  • After each completed reply, a short line reminds you the option exists — controlled by the same per-surface toggles, so switching a surface off makes it fully silent.
Scores travel. A vote cast on any surface appears live on the rating bar of the chat you have open, and reopened conversations show every score they’ve earned. When the same turn ends up voted on twice, the most recent vote wins — a re-vote from your phone can’t be overwritten by a stale copy the desktop was still holding. In-app clicks paint instantly and roll themselves back on the rare failed write.
Scoring is optional everywhere. Unscored turns are still reviewed — just without your verdict. If you only ever score the terrible and the excellent turns, that’s already a strong signal.

​
The Monthly Deep Reflection

Nightly accumulation has a failure mode: self-taught bad habits. Once a month (the 1st, one hour after the nightly run), a deep reflection takes the opposite stance and attacks what the nights built up — stale rules, sweeping conclusions drawn from one incident, entries the evidence has turned against. It audits the playbook and the five knowledge files, reading up to a month of reflection day files as evidence, and rewrites only what changed — knowledge rewrites keep a .bak beside the original, same as the playbook. It skips itself until reflection has actually produced something to audit.

​
Schedule and Settings

Everything lives in Settings → Knowledge → Reflection (the Knowledge page now has two tabs: Compaction and Reflection):
  • Nightly reflection — the hour it runs (default 3 AM).
  • Review after quiet for — how long a conversation must sit untouched before it counts as settled: 1, 2, 3, 6, 12 (default), 24, or 48 hours.
  • Turn scoring — three switches: In-app, Telegram, WhatsApp (all on by default).
  • Run now buttons for both the nightly review and the deep reflection, plus last-run cards showing when each ran, how long it took, the tokens in and out, and what it produced.
The same settings persist in config.json:
Reflection is core — there is no off switch, only the hour. Like compaction, it runs through the brainstem’s scheduler: a laptop asleep at 3 AM simply runs its review on the next launch (missed fires are swept 90 seconds after startup and every 30 minutes after that). Both runs also appear on the Automations page as “Nightly reflection” and “Deep reflection”.

​
Reflection vs. the Other Learning Loops