Skip to main content
Learnings are durable pieces of organization knowledge (conventions, gotchas, preferred approaches) that in-workspace agents fetch at task time instead of relearning every run. Unlike prompt injection, nothing is force-fed into context: the agent queries for relevant learnings when it needs them, and only matching entries are returned.
Learnings is rolling out behind the learnings feature flag. If you don’t see the Learnings tab in your environment settings, contact your Replicas account team to have the flag enabled for your organization.

​
How delivery works

Agents fetch Learnings once at the start of each distinct task, not on every turn; if the task changes, they fetch again. Each learning is delivered one of two ways:
  • Always-apply. Returned on every agent fetch in scope. Use sparingly for universal rules.
  • Trigger-matched. Carries a short “when to use” trigger; a cheap model matches the agent’s task query against triggers and returns only the entries that genuinely apply. If matching is unavailable, the fetch degrades to always-apply entries rather than failing.

​
Scope

A learning is either org-global (applies to every environment) or scoped to a single environment. Manage entries from Dashboard → Environments → the environment’s Learnings tab; use the Global environment for org-wide entries.

​
Human-approval workflow

Agents never write to shared knowledge directly. When an agent proposes an addition, update, or deletion (via replicas learnings), it creates a proposal that a human reviews in the workspace provenance pane. On approval (with optional inline edits), the proposal is applied to the learnings store; declined proposals change nothing. Secrets and credentials are rejected at proposal time and never stored.

​
API access

Learnings are managed through the dashboard and the in-workspace replicas learnings CLI only. The dashboard endpoints are dashboard-session (JWT) authenticated and the agent endpoints require in-workspace agent auth. Unlike environment resources, they are intentionally not exposed on the public API-key surface and are absent from the API reference. This may change once the feature is generally available.