# Amnesia (skipped tools)

An agent can quietly stop using a tool it normally relies on. It keeps
running and keeps answering, but without the search, the lookup, or the
check that made its answers good. Nothing errors, so nothing tells you.
Amnesia is the pre-defined alert that finds these runs: it learns which tools
each of your agents usually calls, and raises a finding when a finished
session skips or loses one of them.

Its findings are filed under the failure mode id `CF-16`. See
[Failure modes](/rius/platform/alerts/failure-modes).

Every workspace with alerting has an **Amnesia** alert, enabled by default.
Like the other [pre-defined alerts](/rius/platform/alerts/pre-defined), you
set its time window on its details page and attach
[channels](/rius/platform/alerts/channels) to it.

## What it finds

Amnesia compares each finished session with what the same agent did in its
own earlier sessions. It reports two rules.

| Rule | Fires when | Severity |
| --- | --- | --- |
| A: offered but never called | The tool was in the agent's tool list on the session's LLM calls, the session never called it, and the agent uses the tool in at least 80% of its sessions | Warning. Critical when the agent uses the tool in at least 95% of its sessions |
| B: missing from the tool list | The agent usually has the tool, this session's tool list did not include it, and a later session has it again | Warning |

A third result, the toolset change, is information only. See
[A removed tool is not a finding](#a-removed-tool-is-not-a-finding).

### Rule A: offered but never called

Fires when the model could have called a tool and never did, in a session
where the agent normally does. For example, a support agent calls
`search_docs` in 92% of its sessions. One session offers `search_docs` on
every call and never uses it, so the agent answered from memory.

The finding names the tool and gives the numbers behind it: how many LLM
calls offered the tool, and the share of sessions that use it. For example:

```text
`search_docs` was offered on 6 LLM calls but never called; this agent uses it in 92% of 41 sessions over the last 7 days
```

### Rule B: missing from the tool list

Fires when a tool the agent usually has was not in the session's tool list at
all, and it shows up again in a later session. The usual cause is a race: the
tool comes from something that connects at startup, such as an MCP server,
and the session began before that connection finished. The agent cannot call
a tool it was never given, so the session ran worse for a reason the agent
had no say in.

Because the tool comes back, Amnesia treats the gap as a race and raises a
warning. It looks up to 24 hours past the end of the session for a later
session that offers the tool again.

### A removed tool is not a finding

If a tool is missing from a session and also missing from the next sessions
of the same agent, you most likely removed it on purpose. Amnesia does not
alert on that. Instead it records an **info** result on the first session
without the tool, such as:

```text
`search_docs` was removed from this agent's toolset since 2026-11-10T14:03:00Z
```

It needs the next four sessions with tool lists to lack the tool too, and
those sessions must close within a day of the first one. The info
result appears on the run, never alerts anyone, and is not counted as a
silent failure. If the tool does come back, the result is the rule B warning
instead.

## What it ignores

Amnesia stays quiet when it cannot judge fairly:

- **Short or errored sessions.** A session needs at least two LLM calls, since
a one-call session has no room to use a tool. A session whose root run
ended in error is skipped too, because it may have stopped early.
- **New tools.** A tool added to an agent's list in the last 48 hours is not
judged, so the agent has time to pick it up.
- **Agents with too few sessions.** Until an agent has 20 other sessions in
the baseline, none of its sessions are judged.
- **Tools the agent rarely uses.** A tool counts as one the agent relies on
only when the agent uses it in at least 80% of its sessions, so a tool it
calls about as often as not is never a finding.
- **Sessions without tool lists.** A session whose LLM calls carry no tool
definitions is unknown, not judged, and does not count as a removal.

The baseline is fixed: the agent's own sessions over the last 7 days, with the
session being judged left out of it. The 80% usage, the 95% critical level, the
20-session minimum and the 7 days do not change per workspace. Two agents
never share a baseline, so a tool that one agent uses all the time does not
make another agent look wrong.

Findings are built from finished sessions: Amnesia judges sessions that
closed within the time window, 60 minutes by default. One agent produces at
most 20 findings for each of the two rules per check, with the strongest
first, so a broken toolset does not raise one finding per tool.

A new workspace, or one that has only just started sending tool
definitions, waits before the first finding. The baseline needs 7 days of
history and at least 20 sessions per agent. Expect no Amnesia
findings during the first week, even if an agent is already skipping tools.

## Read a finding

Findings from Amnesia appear on the Alerts page like the other detectors, one
firing per agent and tool. Open one to see:

- The **tool** and the **agent** it concerns.
- How many **runs** were affected and when you first and last saw it.
- The reason line for the newest run, with the usage share and the number of
sessions it was measured on.

Open a run to see the session itself. To judge a finding, compare it with the
[Tool usage panel](#see-tool-usage-per-agent) of the same agent: it shows how
often the agent uses the tool and lists recent sessions that skipped it.

## See tool usage per agent

The page of each agent has a **Tool usage** panel. For every tool the agent
was offered over the last 7 days it shows:

- The usage rate, such as "39 of 41 sessions, 95%".
- When the tool was last offered.
- Up to five sessions that were offered the tool and skipped it. Each links
to the agent's traces filtered to that session.

Tools the agent never called sit last, muted and labelled "never used". The
panel always covers 7 days, whatever time window the page uses. If the agent
sent no tool definitions in those 7 days, it says so and stays empty.

![The Tool usage panel of an agent: for each tool the share of sessions that called it, when it was last offered, and the sessions that skipped it](/assets/ui_agent_tool_usage.png)

Use it to check a finding before you act on it, and to spot tools your agent
never uses even though you offer them on every call, which cost tokens and
add noise to the model's choices.

## Tune or turn off Amnesia

Open the **Amnesia** alert from the Alerts page to change its time window.

| Setting | Default | Meaning |
| --- | --- | --- |
| Time window | 60 minutes | How far back closed sessions are judged. From 5 minutes to 24 hours |

How much usage counts as reliance, when a finding is critical, and how much
history sets the baseline are fixed, so there is nothing else to tune. See
[What it finds](#what-it-finds) and [What it ignores](#what-it-ignores) for
the values.

To turn Amnesia off, disable the alert on its details page. Disabling it
stops evaluation, so it raises no new findings; its open firings stay as they
are until you enable it again. You cannot delete a pre-defined alert.

## What to send

Amnesia can only judge what your agents report. A quiet alert can mean healthy
agents or missing inputs, so compare this list with what your spans carry.

| Amnesia needs | Which means |
| --- | --- |
| Tool definitions on LLM spans | The list of tools each LLM call was given, as `gen_ai.tool.definitions`. The Rius SDKs send it while content capture is on |
| Tool call spans | A span with a tool name for each call the agent makes, so Amnesia can tell which tools were used |
| An agent name | `gen_ai.agent.name` on the agent spans, so the baseline is per agent |
| A session id | Optional. Without one, each trace counts as its own session |

Tool definitions are content, so they follow your privacy settings. With
[`capture_content=False`](/rius/sdk/advanced#capture_content-the-content-switch)
the SDKs do not send them and Amnesia has nothing to compare. Tool names on
tool call spans still flow.

If an agent sets no name, Rius falls back to the service name and puts every
agent of that service into one baseline. A tool that one of them uses would
then look expected for all of them. Set the agent name so each agent is
measured against itself. The console does not flag a missing agent name on its
own, so check it yourself.

The console tells you when Amnesia cannot run. In the detector coverage for
the workspace, Amnesia shows one of these reasons while it is not ready:

| Reason | What to do |
| --- | --- |
| No LLM spans carry tool definitions | Turn on content capture in the SDK so tool definitions are sent |
| Tool definitions are present but no tool spans were seen | Check that your tool calls are traced |

## Use it from your coding agent

The [Rius MCP server](/rius/mcp) has a tool for the same data as the Tool
usage panel, `list_agent_tool_usage`. Ask your coding agent which tools an
agent skips, or call it with the agent name and the workspace id from
`list_workspaces`:

| Argument | Type | Default | Description |
| --- | --- | --- | --- |
| `workspaceID` | string | required | The workspace id, from `list_workspaces`. |
| `agentName` | string | required | The exact agent name, from `list_agents`. |
| `hours` | integer | `168` | How far back to look, from 1 to 720. |

> Which tools does checkout-agent skip?

```json
{
    "name": "list_agent_tool_usage",
    "arguments": { "workspaceID": "<workspace id>", "agentName": "checkout-agent" }
}
```

It returns the number of sessions with tool definitions, and for each tool:

| Field | Meaning |
| --- | --- |
| `tool`, `display_name` | The tool's name |
| `sessions_offered`, `sessions_called` | Sessions that were offered the tool, and sessions that called it |
| `usage_pct` | `sessions_called` as a share of `sessions_offered`, to one decimal |
| `never_used` | `true` when no session called the tool |
| `first_offered_at`, `last_offered_at` | When the tool was first and last in a tool list |
| `skipped_session_ids` | Up to five of the newest sessions that were offered the tool and skipped it |

Used tools come first, highest usage first. Tools never used come last. An
agent with no tool definitions in the period returns an empty list. Example
sessions are looked up for the first 30 tools that have a skipping session, so
later tools show an empty list.

To find the findings themselves, use `list_silent_failures` with
`failure_mode` set to `CF-16`. See the [tool reference](/rius/mcp/tools#list_silent_failures).

## Next steps

- [Pre-defined alerts](/rius/platform/alerts/pre-defined): the other detectors
and how knobs and channels work.
- [Context rot](/rius/platform/alerts/context-rot): the detector for full,
compacted, and truncated contexts.
- [Distribution channels](/rius/platform/alerts/channels): route findings to
webhook, email, or Slack.
- [Privacy controls](/rius/sdk/advanced#privacy-controls): decide what content
leaves your process.
