--- title: Self-improvement description: Runs that compound. Agents recall the organization's know-how at the start, propose what they learned at the end, and nothing publishes without a human accepting it. sidebar: order: 7 --- Most agent platforms are amnesiac: every run starts from zero and whatever it figured out evaporates at the end. UseAgent closes the loop. A run **starts with what the organization already knows** and **finishes by proposing what it learned**, with a human gate between proposal and publication. ## The substrate agents draw from | Store | What it is | | --- | --- | | **Skills** | Versioned, reusable procedures, selected semantically and activated by exact id. | | **Playbooks** | Skills whose content is a repeatable end-to-end operating procedure. | | **Knowledge** | Organization-scoped source material an agent can search and read. | | **Wiki** | Published documents generated from repository or organization knowledge, retrievable by agents later. | | **Memory** | Optional scoped facts retained across sessions, independently disablable. | Skills and playbooks share one immutable substrate, and 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. Recall is visible, not ambient: what was recalled renders in the timeline as context markers (memory, knowledge, skill, playbook, rule), so a reviewer can see exactly what shaped a turn. ## The loop, end to end 1. **Recall at start** A run recalls organization memory when it begins, through the same gateway tool surface everything else uses. 2. **Capture at finish** Capture is salience-gated and enriched with structured evidence, including the ordered tool trace of what actually worked, redacted before it goes anywhere. 3. **Propose, do not publish** High-value runs propose knowledge drafts; accepted drafts that cluster propose skill revisions whose procedure is assembled from the real traces. 4. **A human accepts or dismisses** An org admin reviews each proposal at /learnings. Accepting publishes to the stores and re-projects into retrieval, so the learning becomes findable in future runs. Nothing auto-publishes. That last property is the point: **there is no unattended write path into the knowledge stores.** The machinery is complete, and throughput is deliberately gated on human review. See [Learnings review](/product/learnings) and [Knowledge and learning](/concepts/knowledge-and-learning). ## Import what you already wrote Organizations rarely start empty. UseAgent scans a GitHub repository for `SKILL.md` files and imports selected paths as versioned skills: - Discovery runs over a server-side shallow clone. - Imports are **pinned to a commit** and **idempotent by source**, so re-running an import never duplicates. - Periodic resync is opt-in and operator-timed, disabled unless an interval is configured (`backend/src/skills/resync.ts`). ## Memory you can audit Memory is the store most platforms make invisible. UseAgent gives it a full surface at `/memory`: What is stored in a pool, with correction and deletion. What each run wrote, and its delivery state through the durable outbox. The recall ledger: which memory was used in which runs. The safety properties are structural: the organization pool is shared while the personal pool is private and **fails closed** without an authenticated user; a run's memory scope is server-persisted, never taken from the sandbox at recall time; and memory is config-gated, so a memory outage degrades to no-op rather than failing runs. See [Memory hub](/product/memory-hub). :::note **Compounding is earned, not claimed.** The loop improves an organization only as fast as its reviewers accept drafts. That is a deliberate trade: slower compounding, zero unreviewed contamination of the stores every future run reads from. :::