> ## Documentation Index
> Fetch the complete documentation index at: https://docs.joystream.ai/llms.txt
> Use this file to discover all available pages before exploring further.

# Evidence, memory, and metrics

> What counts as proof that a capability worked, what an agent is allowed to learn, and how outcomes are measured.

## Evidence

Evidence is proof that the intended capability outcome occurred.

JoyStream does not stop at technical execution. A successful tool call is not necessarily a successful capability.

| Technical success          | Capability success                                                                            |
| -------------------------- | --------------------------------------------------------------------------------------------- |
| The email API returned 200 | The correct employee received the correct onboarding information before the required deadline |

Evidence may include tool receipts, created artifacts, before-and-after state, approvals, validation results, screenshots, structured outputs, semantic evaluation, human confirmation, business metrics, and policy compliance.

It exists at three levels:

```mermaid theme={null}
flowchart TD
    A["<b>Story evidence</b><br/>Did this work item succeed?"]
    --> B["<b>Run evidence</b><br/>Did this execution complete correctly?"]
    --> C["<b>Capability evidence</b><br/>Is the capability reliably delivering<br/>its intended organizational outcome?"]
```

<Note>
  A run is not successful merely because execution stopped without an error. The intended outcome must be demonstrated.
</Note>

Evidence is what allows an agent to earn trust — and it is what a person reviews when deciding whether to promote an agent to Pilot or Live.

## Memory

Memory is the accumulated context and learning produced through the operation of a capability: past decisions, organizational preferences, exceptions, stakeholder context, feedback, historical outcomes, successful patterns, failed approaches, policy interpretations, and reusable knowledge.

Memory belongs primarily to the capability — not to one conversation, one model, one agent version, or one run. That is what lets a capability get better across model changes and staff changes.

Different kinds of memory have different scopes: story, run, agent, capability, workspace, and organizational knowledge.

Memory should be attributable, permissioned, inspectable, correctable, and version-aware.

### Memory cannot silently override governance

<Warning>
  Memory informs execution. It cannot silently rewrite policies, boundaries, success criteria, or organizational intent.
</Warning>

Learned behavior becomes authoritative only through review. If an agent should behave differently, that change goes through a new capability spec version — not through something the agent quietly picked up on a Tuesday.

## Metrics

Metrics measure whether the capability is creating organizational value. They come in three layers, and only the third is the point.

| Layer           | Examples                                                                                                                                                    |
| --------------- | ----------------------------------------------------------------------------------------------------------------------------------------------------------- |
| **Technical**   | Latency, cost, failure rate, retry rate, tool reliability, model usage                                                                                      |
| **Operational** | Completion rate, exception rate, human intervention, queue time, approval time, throughput                                                                  |
| **Capability**  | Time to onboard, percentage ready before day one, lead qualification accuracy, renewal risk detected, incidents resolved, financial close completed on time |

JoyStream optimizes for capability outcomes, not agent activity. An agent that runs constantly and delivers nothing is not succeeding.

## Human accountability stays visible

Every capability has identifiable human ownership. AI agents execute and collaborate. People remain accountable for intent, access, governance, exceptions, promotion, and organizational consequences.

## Next

<Columns cols={2}>
  <Card title="Workspaces and organizations" href="/concepts/workspaces-and-organizations" icon="building">
    Where ownership and access live.
  </Card>

  <Card title="Review runs and tickets" href="/guides/review-runs-and-tickets" icon="clipboard-list">
    Inspect what a run actually produced.
  </Card>
</Columns>