CTXORA Continuous Learning

CTXORA learning is deliberate and memory-based; it does not require background hooks.

Skill-routing learning is separate from durable memory. Keep the route_id returned by route_skills or prepare_context in task state. After completing a task and verifying validation, automatically perform self-evaluation against the 5 quality axes (accuracy, completeness, clarity, actionability, conciseness) and call skill_feedback with success and the skills actually used; use failure only when the failed result is attributable to the guidance, and corrected when the user replaces the routed skill. Do not ask the user to submit routine feedback manually. CTXORA stores only a task fingerprint, matched catalog terms, bounded counters, a bounded outcome ledger, and decayed associations; it never stores raw tasks or transcripts.

CTXORA does not infer task success from Git changes or merely from sharing a worktree. Without skill_feedback, a routed task remains pending and must not affect positive reranking. Use skill_learning_status or the local Console to inspect completed outcomes, pending routes, and current boost/penalty signals.

Inspect connected tool schemas and resolve the registered workspace ID before using memory_search or memory_save. If tools are unavailable, present a proposed lesson and explicitly report that it was not saved. Do not write directly into memory databases or ECC vaults. Do not capture raw session transcripts or install observers as part of this skill.

Before saving, search existing memory for the same rule. Save only knowledge supported by repository evidence, a passing validation, an explicit user decision, or a reproduced incident.

Choose one memory tier:

• semantic: stable project facts, terminology, conventions, and decisions.
• procedural: repeatable workflows, diagnostics, and repair sequences.
• episodic: a specific incident, handoff, experiment, or one-time decision context.

Use workspace scope by default. Use a broader scope only when the same lesson has been independently verified across projects. Store one atomic lesson per key, include the evidence or validation in the value, set source accurately, and lower confidence when evidence is incomplete.