ETM

ETM stands for Episodic Task Memory. Nova uses episodic task memory for task-shaped episodes: what the goal was, which context applied, what was checked, what remained open, and when the task was truly complete.

ETM is Nova’s task memory. It lets agents recognize recurring task types, start concrete task runs, keep progress visible, and resume work without pretending that an unfinished job is done.

ETM is different from PKS. PKS stores reusable website knowledge; ETM stores task-shaped knowledge. PKS can say how a known website pattern behaves. ETM can say which task is running, which units belong to it, and what evidence is still needed before completion.

ETM tools are mostly agent-facing. Users see the effect in resumable work, clearer automation, and fewer premature completions, while agents and integrators use the task-memory tools directly.

In short

• ETM stores task knowledge, not general website behavior.
• Task profiles describe reusable task types; task instances describe concrete runs.
• Work units make large tasks measurable instead of relying on a narrative “done”.
• Completion can require all units, a threshold, or an exploratory minimum.
• Task URL Coverage belongs to ETM and checks URL coverage when a task claims exhaustive completion.
• ETM can turn repeated successful runs into stronger guidance for later work.

What ETM stores

ETM stores the information an agent needs to understand work as a task, not just as a series of browser actions.

1. Task profiles A profile describes a recurring task type: goal, domain or platform context, stable guidance, required checks, known exceptions, and a completion rule.
2. Task instances An instance is one concrete run. It may start from a profile or from an ad-hoc task, then freezes the effective context so later profile changes do not rewrite the current run.
3. Work units Work units are bounded items that need attention, such as pages, URLs, findings, records, checks, or other parts of the task.
4. Progress and resume state ETM can keep checked, remaining, blocked, and failed work visible so an agent can resume instead of guessing what happened earlier.
5. Completion conditions A completion rule says what “done” means for this task: all units processed, a threshold reached, or enough exploratory evidence collected.
6. Guidance and learning Repeated runs can leave guidance behind. Useful guidance can later become part of a profile when enough evidence supports it.

When this matters

ETM belongs to tasks that are larger than one quick interaction or likely to return later. Typical examples are website reviews, repeated quality checks, scheduled automation, platform mapping, research across multiple sources, or any run where an agent must prove what was covered before it can stop.

Task model for agents

For an agent, ETM answers four questions: which reusable task is this, which concrete run is active, what work remains, and what evidence is required before completion?

Profile and match

A profile gives the agent a known task shape. Matching uses task description, task type, domain, platform, target URL, and current scope to find likely profiles.

Instance and context

An instance is the current run. It carries a frozen effective context, profile revision, target URL, current scope, progress, and resume state.

Task awareness

A compact agent-facing payload that can show a profile match, active instance, progress, completion rule, pending checks, and resume context.

Work units

Units move through discovered, checked, excluded, blocked, or failed states. Checked and excluded units count as done; blocked and failed units reject exhaustive completion and should be repaired before a final claim.

Discovery state

Discovery starts unknown, can become partial, and can be frozen when the agent believes the task frontier is complete. Exhaustive completion requires a frozen frontier.

Mandatory checks

Required checks must be satisfied, waived, or marked not applicable before completion should be accepted. Pending or failed checks block completion.

Completion condition

The rule that defines done. ETM supports exhaustive, threshold, and exploratory task shapes with explicit stop metrics.

Guidance logs

Agents can log task guidance, inspect repeated hints, and promote useful guidance into the reusable profile when it proves useful.

How agents use ETM information

For agents, ETM is a structured task surface: discovery hints, profile matches, active instances, progress counters, completion rules, revision guards, verification contracts, and coverage warnings tell the agent how to proceed.

1. Search or match the task Before starting task-shaped work, the agent should search or match existing profiles. A match can return task awareness and reusable guidance.
2. Start the right instance The agent creates a task instance from a profile or ad-hoc context. This instance becomes the stable anchor for progress, current scope, and later completion.
3. Report progress with revision guards Progress updates use the current instance revision and an idempotency key. The agent reports discovered units, unit status changes, findings, mandatory-check updates, and resume-state deltas.
4. Freeze discovery for exhaustive work For exhaustive tasks, the agent freezes discovery only after it has found the relevant frontier. A frozen empty frontier or open units can still reject completion.
5. Check completion before claiming done ETM evaluates mandatory checks, unit state, completion mode, revision conflicts, verification evidence, and coverage state. A failed check means the task remains unfinished.
6. Resume from task awareness When a run continues later, the agent reads the active instance, recent events, pending units, and resume state instead of reconstructing the task from memory.
7. Promote useful guidance After repeated runs, guidance logs and match telemetry can show which instructions should become stable profile guidance.

When ETM state applies

ETM state is about the current task run, not permission to click on a page. An agent may only claim completion when the active instance and the current evidence satisfy the task contract.

Completion and failure conditions

These are the contract checks that decide whether ETM accepts a task as complete. They are intentionally stricter than an agent saying that work feels done.

Task URL Coverage

TUC stands for Task URL Coverage. It belongs inside ETM and is used for tasks with an exhaustive completion claim. Agents can create URL units with nova.task_instance_create, gather server-trusted coverage evidence with nova.coverage_scan, and reconcile open coverage with nova.task_instance_reconcile_coverage. These are agent-facing tool identifiers, not normal UI controls for people.

Agent interpretation example

The example shows ETM as a task contract: the agent creates a run, reports progress, and may only complete when the completion checks allow it.

Task match

{
  "tool": "nova.task_match",
  "arguments": {
    "taskDescription": "Review all product pages for broken prices",
    "targetUrl": "https://example.com/products",
    "currentScope": {
      "route": "/products",
      "locale": "de-DE"
    }
  }
}

Progress update

{
  "tool": "nova.task_instance_progress",
  "arguments": {
    "instanceId": "task_123",
    "expectedInstanceRev": 4,
    "clientEventId": "price-check-page-18",
    "unitUpdates": [
      { "unitKey": "product-18", "status": "checked" }
    ],
    "mandatoryCheckUpdates": [
      { "checkId": "price-visible", "state": "satisfied", "evidenceRef": "product-18" }
    ]
  }
}

Agent interpretation

{
  "treatAs": "active task contract",
  "beforeCompletion": "read the instance, check remainingUnits and pendingMandatoryChecks",
  "mayComplete": "only when completionAllowed is true and revision is current",
  "onRejection": "use currentState.reason to continue, repair, exclude, or gather evidence"
}

MCP contract

This is the deterministic layer below the explanation. It lists ETM fields that agents and integrators should treat as contract signals rather than free prose.

Completion rule: No ETM completion may be claimed without current task state, revision, required checks, and evidence satisfying the active completion condition.