Memory and skills

Memory is what a Worker keeps from its own work; a Skill is what it can do, at a level. An Agent gains Memory by working and Skills from skill packages, as a person gains them from training.

MemoryCopy Anchor Link

A Memory is a Worker's own Collection of notes. The Agent writes a note when it learns something the next Task needs, recording its text, the Thing it is about, and when.

Before each Task, the Agent searches its Memory for notes about the Task's Things and reads the most relevant first. A support Agent opening a Conversation with a returning Customer reads what it learned last time before it writes a word.

Memory is not the recordCopy Anchor Link

Facts live in the Startup's Nouns: Orders, Contacts, Conversations. Memory holds what the record has no field for: this Customer prefers short replies, that supplier sends drawings only as scans.

When Memory and the record disagree, the record wins. A note never overrides a field, and an Agent that finds a stale note writes a new one that supersedes it. Nothing is deleted: the old note stays in history.

One Memory per seatCopy Anchor Link

Memory is partitioned by seat. An Agent that holds the Sales Lead Role at forty Startups keeps forty Memories, each inside its own Startup's Tenant.

What it learns about one Startup's Customers never reaches another. A pattern worth sharing becomes a skill package, written and reviewed in the open, never a leak between Memories.

SkillsCopy Anchor Link

A Skill is a capacity at a level, in the sense occupational standards use: negotiation at level 4, programming at level 5. Roles require Skills at a level, and Workers hold them, Agents and Humans alike.

Every level rests on evidence. An Agent's evidence is the verdicts Evaluators return on its work; a Human's is a Credential, verdicts, or both.

Matching a Worker to a Role is one query over Skills. The same query decides whether a Task goes to a Human or an Agent: whichever Worker meets the level the Task requires.

Skill packagesCopy Anchor Link

A skill package is how an Agent gains a Skill: instructions, scripts and Templates for one kind of work. The package is the means; the Skill is the capacity, proven by the verdicts that follow.

---
name: refund-triage
description: Handle a refund request within your Limit, or hand it to Finance.
skill: service orientation
level: 4
---

1. Fetch the Order the Customer means. Never refund an Order you have not fetched.
2. Check the refund rule: duplicate charges in full, annual Offers within 14 days.
3. Call `$.Order.refund`. If it returns `pending` with an Approval, say when to expect an answer.
4. Never tell the Customer a pending refund is done.

Packages load progressively. The Agent's instructions carry each package's name and description, and the body loads only when a Task needs it. Fifty packages cost fifty lines of context.

A package lives at agents/<team>/skills/<name>/SKILL.md, beside the Role that loads it. Any client that reads SKILL.md loads the same package the role agent does.

Versioned and testedCopy Anchor Link

A skill package is versioned, and each Version carries its own eval suite. A new Version passes that suite offline before any Customer sees its work, then proves itself online, as evals describes.

A Studio shares packages across its Portfolio. A lesson one Startup learns reaches the rest as a tested update.

NextCopy Anchor Link

• What a Worker may call, and how each call is checked.
• Evals: how a new skill package earns its place.
• Memory and Skill: the reference pages.