Continuous Learning

Continuous Learning

Turn demonstrated outcomes into small, durable improvements without overfitting to a single incident.

Identify a Candidate Lesson

A useful lesson has evidence and future value. Look for:

• a repeated failure with a confirmed cause;
• a non-obvious project invariant;
• a workflow that consistently reduced errors or effort;
• a tool or command whose exact usage is hard to rediscover;
• a decision boundary that prevented misrouting or unsafe action.

Do not preserve guesses, one-off preferences, generic advice, sensitive data, or rules already enforced by the host environment.

Generalize Carefully

Record the triggering situation, observed evidence, lesson, scope, exceptions, and verification. Keep a project-specific lesson in project guidance. Put a reusable workflow in a skill only when it applies across realistic tasks and changes agent behavior.

Prefer a narrow correction over adding an absolute rule. A past failure does not prove every similar situation requires the same response.

Validate

Before proposing persistence:

1. confirm the outcome and causal evidence;
2. check for an existing authoritative instruction;
3. identify conflicts or duplication;
4. test the lesson against at least one different but realistic scenario;
5. choose the smallest appropriate destination.

Persist Only with Authorization

Show the proposed update and obtain any authorization required by the host environment. Never silently edit shared instructions, install skills, or store personal context.

After updating, report the destination, exact lesson, and evidence. Add a review date when the lesson depends on a tool, API, policy, or other drift-prone fact.