Skills your agents learn

Most AI tools start from zero every morning. Unyo doesn't. When one of your agents completes a run that contains a real, repeatable procedure, an overnight worker writes that procedure down — and shows it to you. If you keep it, the agent starts applying it to similar work from then on.

A skill is not a memory of what you said. It's a method: the sequence of moves that made a piece of work come out right. Your Neural Core remembers your facts; skills remember your way of working.

How it works, end to end

1. You work normally. Ask Blake for a KPI breakdown, Lucy for a comparison article, Tyron for a resized visual. No flag to set, no "teach me" mode.
2. Overnight, a worker looks back. Once a day it scans your recent completed runs and examines at most eight of them.
3. It distils at most one procedure per run. Usually zero. Silence is the default — the worker is instructed to return nothing when in doubt, and it takes that instruction seriously.
4. You get one line, in your chat. A small card appears above the composer: "🧠 Blake learned something", plus a one-sentence summary in your own language.
5. You decide. Keep activates it. Forget archives it. Later leaves it in the queue and asks you again another time.
6. Kept skills get injected. On every future run, the three best-matching active skills for that agent and that goal are added to the prompt — before the agent starts working.

What the distiller can actually see

This is the part worth reading closely, because it's the part most "AI that learns from you" claims quietly skip.

The worker that writes your skills receives exactly three things:

Then it checks its own work. Anything it writes is run through a pattern check for email addresses, URLs and long digit strings — if a skill leaks one, the skill is rejected outright rather than shown to you.

What it refuses to learn

The worker skips far more than it keeps, on purpose:

• Runs with fewer than two tool calls. One tool call is not a procedure.
• Runs with no real goal — anything under 15 characters.
• Anything that names your data. Names, emails, companies, product names, prices and dates are banned by instruction and by a check that runs afterwards.
• Transient failures. "The store is down", "that endpoint is broken" — a bad afternoon is not a skill. This is blocked by prompt and by a second pattern check on the output.
• Duplicates, caught three separate ways: by instruction, by a similarity check against your active skills, and by a word-overlap check against your pending candidates (so two similar days can't spawn twins).

If none of the runs it examined contain a procedure worth keeping, it writes nothing and you see no card. That is the normal outcome.

Which skills get used, and when

When you send a goal, Unyo looks up your top three matching active skills for that agent and injects them into the run. The ranking is a weighted blend:

Skills scoped to a single agent and skills scoped to all agents both compete for those three slots. A skill that matches nothing is simply not used — there is no penalty for having many.

Every agent can do this. skill.list returns your active skills for that agent (plus the ones scoped to all agents), ranked by trust; the agent can then pull up any one in full. It's the escape hatch for when automatic matching misses.

Keeping, forgetting, and retiring

• Keep flips the candidate to active and raises its trust to at least the halfway mark. From that moment it competes for the three recall slots.
• Forget archives it. Archived skills are never recalled and never shown again.
• Later decides nothing. The card comes back.

Unyo will only ever hold five pending candidates for you at once. Once the queue is full, the overnight worker stops proposing and waits — so a week away doesn't come back to 40 approvals.

After that, a job runs every hour and tallies outcomes against the skills that were recalled into each run. A distilled skill that has been recalled at least ten times, and whose runs failed more than twice as often as they succeeded, is archived automatically.

Auto-archiving only ever touches skills the system distilled by itself — including ones you kept. A skill you wrote yourself is never retired by the tally.

Where skills live

Skills are stored per account, protected by the same row-level security as everything else in Unyo: you can read and delete your own, and nothing else can read them. They hold instructions only — no tokens, no keys, no personal data by construction, since the distiller was never shown any. They cascade-delete with your account like every other table.

Tips

• The best skills come from your boring work. A one-off creative request rarely contains a procedure. A report you run the same way every week almost always does.
• Read the card before you tap. The one-sentence summary is written in your language specifically so the decision takes two seconds.
• Forget freely. Archiving costs nothing, and a wrong skill that stays active competes for a recall slot a good one could have used.
• Keep your interface language stable. Skills are written in it, and matched against goals you type — mixing languages weakens matching.
• If an agent seems to ignore a skill, say so. Ask it to list its skills and open the one you mean. Recall missing a skill is a wording problem, not a broken skill.

Troubleshooting

I've been using Unyo for days and no agent has learned anything. Expected, and usually correct. The worker examines a small number of recent completed runs per pass and rejects anything that isn't a genuinely repeatable, non-obvious procedure. Runs with fewer than two tool calls are skipped outright. Short chats and one-shot questions produce nothing — by design.

I tapped Keep but the agent didn't use it on the next run. Being active makes a skill eligible, not guaranteed. Only the top three matches for that specific goal are injected, and matching is on wording. Ask the agent to list its skills and open the relevant one directly.

The same card keeps coming back. You tapped Later, which deliberately decides nothing. Use Keep or Forget to clear it.

A skill went stale after I changed how I work. Tap Forget the next time you see it referenced. Recency is 20% of the ranking, so a skill you stop triggering fades on its own — but one that's still being pulled into runs won't, so retire it explicitly. The next run of the new way of working is a fresh distillation candidate.

FAQ