Learning

Learning in Nova does not mean that an agent keeps arbitrary memories. Nova separates observed work, learning candidates, reviewed website knowledge, current operational state, and concrete task runs so later guidance stays useful and checkable.

This overview is the map for Nova’s learning area. Tool Observation Bus (TOB) records observed tool evidence. Lightweight Candidate Journal (LCJ) keeps learning evidence. Agent Learning Pipeline (ALP) reviews candidates and promotes the useful ones. Closed-Loop System (CLS) verifies action outcomes. Phenomenological Knowledge Store (PKS) stores curated website phenomena. Operational Knowledge (OK) keeps current service state. Episodic Task Memory (ETM) remembers task runs, progress, and completion conditions.

TUC stands for Task URL Coverage. TUC belongs inside ETM. It checks whether relevant URLs or work units were actually covered before a task with an exhaustive completion claim is accepted as complete.

Adjacent memory surfaces stay separate from the learning pillars. Browser Memory keeps opt-in user context for a domain, Operator Notes keep reusable cross-session notes, and Evidence Verification Mode (EVM) guides research-style claim checking through task instructions.

In short

• Learning starts with observed evidence, not with an agent claim.
• Automatic learning can prepare candidates, but candidates are not active trusted knowledge yet.
• ALP decides whether repeated evidence is useful enough to be promoted.
• CLS checks whether an action actually produced the expected result before the next step trusts it.
• PKS, OK, and ETM keep different kinds of knowledge so guidance does not blur together.
• TUC belongs to ETM because it checks coverage for concrete task runs.
• Browser Memory and Operator Notes are adjacent memory surfaces, not replacements for PKS, OK, LCJ, or ETM.

Learning Areas

Each learning area answers a different question. The point is not to store more memory everywhere, but to keep evidence, review, current state, and task progress in the right place.

The Learning Chain

The sequence maps responsibilities. Nova may enter the chain at different points, but each area has a clear job.

1. Observe TOB records what actually happened while an agent used tools: reads, clicks, inputs, navigation, failures, and blockers.
2. Collect LCJ keeps learning evidence and candidates without treating them as trusted knowledge yet.
3. Review ALP scores repeated evidence, explains decisions, and promotes only candidates with enough support.
4. Verify CLS checks whether a browser action really produced the expected outcome before that outcome feeds guidance or task progress.
5. Store PKS and OK separate durable website behavior from current operational facts such as visible state, service capability, or freshness.
6. Use PKS, OK, and ETM return guidance to later agents while still requiring the current page or task state to confirm the match.
7. Prove completion ETM and TUC keep task progress, open units, completion rules, and URL coverage visible before a large task is declared done.
8. Maintain ALP and PKS surface drift, stale knowledge, repeated failures, and candidates that should be patched, demoted, or retired.

Kinds of Learning

Nova uses several learning paths because evidence can be collected automatically, requested deliberately, reviewed semantically, or tied to a concrete task run.

1. Automatic learning When safe actions repeat, selectors keep working, or clear website situations reappear, Nova can prepare candidates automatically. This reduces manual bookkeeping but does not make the knowledge active by itself.
2. Deliberate Learn Mode An agent can ask Nova for learning-oriented instructions through nova.get_instructions and produce a structured PLATFORM_PLAYBOOK. These names are for agents and integrators, not normal user controls in the UI.
3. Semantic learning Recurring situations such as consent dialogs, login walls, or risky repeated actions can become learning opportunities. Agent-facing tools such as nova.learn_resolve_opportunity help decide how to handle them.
4. Task learning ETM keeps task profiles, task instances, progress, guidance, and completion rules. TUC extends this when a task needs evidence that relevant URLs or units were covered.

Where It Belongs

The learning area should stay readable by giving each concept a home.

Website behavior

belongs in PKS when it is reusable and reviewed.

Current service state

belongs in OK when it describes what is true right now.

Task progress

belongs in ETM when it describes a concrete run, open work, or completion.

URL coverage

belongs in ETM through TUC, because coverage is evidence for a task run.

User context for a domain

belongs in Browser Memory when the user wants Nova to remember notes, preferences, or session context across visits.

Reusable operator notes

belong in Operator Notes when a preference, research rule, or workflow fact should apply across tasks instead of one website pattern.

Research claim checking

belongs to EVM task guidance. It can write claim evidence into LCJ, but it is not a separate long-term learning pillar.

Agent awareness and safety hints

can support learning, but they are not learning pillars on their own.

Limits

Learning should make Nova more reliable, not less accountable. A candidate can remain untrusted, stored knowledge can be stale, task completion can be rejected, and sensitive actions still need the normal safety and approval model. The current page, current task state, and current user intent remain decisive.