Example: demonstrate persistent memory and agent learning

Parent: #588

Goal

Add a progressive, feature-focused example showing how CAO carries knowledge across sessions and turns repeated outcomes into reusable agent lessons.

The sample should connect memory and learning as one vertical: store and recall a durable fact, prove injection into a fresh agent context, record outcomes, distill a lesson for a target agent profile, and observe that lesson on a later task.

Scope

• Add examples/memory-learning/ with a README, runnable entry point, profiles, a small deterministic task corpus, and isolated state.
• Run with a temporary CAO_HOME_DIR so the example never reads or modifies the operator's real memory store.
• Part 1: store a project or agent-scoped convention, start a fresh terminal, and prove it can recall or receive the convention through normal memory injection.
• Part 2: enable memory.learning_enabled, record several structured outcomes, and hand off to the retrospector.
• Store one concise lesson in the intended worker profile's agent scope and prove that a later worker can recall or receive it.
• Use a deterministic scorer or observable rubric so the sample demonstrates changed behavior rather than merely listing stored memory.
• Show instruction promotion as a reviewed dry-run. Do not automatically execute cao memory promote --apply.
• Document scope/type selection, feature gates, privacy boundaries, and cleanup.

Acceptance criteria

• All memory and outcome state is isolated under a temporary CAO_HOME_DIR.
• A fact stored in one session is available to a fresh session through a documented CAO memory path.
• Outcome capture and retrospector reads are disabled until learning is explicitly enabled.
• The retrospector stores a 1-2 sentence lesson with an Applies when: trigger in the target worker's agent scope.
• A later worker receives or recalls the lesson and the sample records an observable before/after result.
• Promotion output is reviewable and dry-run by default.
• The sample does not store prompts, transcripts, logs, credentials, or fixture secrets as memories or outcomes.
• Deterministic tests cover memory isolation, outcome capture, target-agent lesson placement, and cleanup; live-provider validation may be separately gated.

Non-goals

• Claiming that learning improves every task family.
• Automatically promoting untrusted agent-authored instructions.
• Demonstrating the memory graph, OKF export/import, or every memory maintenance command.