---
title: Knowledge and learning
description: Skills, playbooks, knowledge, wiki, and memory, plus the human-gated loop that turns a successful run into a reusable procedure.
sidebar:
  order: 7
---

UseAgent gives an agent an organization's accumulated know-how at the start of a
run, and captures what a run figures out at the end. Nothing is published without
a human.

## The knowledge substrate

| Term | What it is |
| --- | --- |
| **Skill** | Versioned, reusable instructions, selected semantically and activated by exact id for the current task. |
| **Playbook** | A skill whose content describes a repeatable end-to-end operating procedure. |
| **Knowledge** | Organization-scoped source material an agent can search and read. Reference data, not executable instructions. |
| **Wiki** | A published document generated from repository or organization knowledge and later retrievable by agents. |
| **Memory** | Optional scoped facts retained across sessions for a user or team. Reference material, and independently disablable. |

Skills and playbooks share one immutable substrate. The turn-prefill catalog is
re-ranked by prompt relevance on every unpinned turn, so the procedures that
match the current ask surface into the page the model sees, rather than a fixed
top-N by usage.

## Skill import

GitHub is the source of an organization's procedures. UseAgent scans a repository
for `SKILL.md` files and imports selected paths as versioned skills:

- Discovery runs over a **server-side shallow clone**, resolving the head with
  `git ls-remote`, because the REST git-data surface is not dependable for org
  installations.
- Imports are **pinned to a commit** and **idempotent by source**, so re-running
  an import does not create duplicates.
- Resync is **opt-in and operator-timed**: a periodic sweep runs only when
  `SKILLS_RESYNC_INTERVAL_MIN` is set to a positive number of minutes, and
  unset means disabled. Each sweep is capped at 200 repositories with 2-second
  pacing between repositories (`backend/src/skills/resync.ts`).

## The learning lane

The self-improvement loop runs on every completed run and is fully human-gated:

1. **Recall at start**

    A run recalls organization memory when it begins.

2. **Capture at finish**

    At the end, capture is salience-gated and enriched with structured evidence,
    including the ordered tool trace of what actually worked.

3. **Propose knowledge drafts**

    High-value runs propose knowledge drafts, carrying the redacted procedure
    trace.

4. **Propose skill revisions**

    Accepted drafts that cluster propose skill revisions whose procedure is
    assembled from the real traces.

5. **A human reviews**

    An org admin accepts or dismisses each proposal at `/learnings`. Nothing
    auto-publishes.

:::note
The machinery is complete, but throughput is intentionally gated on human review.
The loop compounds only once a person accepts the first draft; there is no
unattended write path into the knowledge stores.
:::