memory-adaptation

name	memory-adaptation
description	Use when a coding task should adapt to durable user/project preferences stored in omni-mem; skip for trivial factual/chat requests.
policy_doc_kind	skill
classification	canonical
canonical_owner	self
authority_level	procedural
in_verifier_scope	true
lexical_guard_profile	stale_names,destructive_rollback,branch_policy_live

Memory Adaptation Skill

This skill owns omni-mem preference retrieval and adaptation workflow only. Global and workspace policy own routing, git safety, review requirements, and notification rules.

Use When

• The user asks to adapt style, workflow, or implementation behavior to prior preferences.
• The task is non-trivial and preference-aware behavior can change planning or implementation decisions.
• Prompt contracts, planning gates, or review output require preference retrieval evidence.

Do Not Use When

• The request is a one-off factual answer with no implementation.
• The task is trivial (single-line/simple edit) and preference retrieval adds no value.
• The user explicitly asks to ignore memory for this run.

Workflow

1. Retrieve context with mcp__omni-mem__build_context scoped to cwd.
2. Run targeted mcp__omni-mem__search queries for pref: keys.
3. Resolve active preferences with deterministic order:
   1. explicit user > project > workspace > global
   2. higher confidence
   3. newest timestamp
4. Select and apply a prompt contract from references/PROMPT_CONTRACTS.md.
5. Record implementation evidence fields:
   • memory_retrieval_evidence
   • preferences_applied
   • skill_trigger_eval_results
   • prompt_contract_used
   • frontend_roundtrip_evidence (frontend scope only)
6. Save durable updates with save_preference (stable preference) or save_memory (episodic lesson).

Degraded Memory Handling

• If omni-mem returns a structured runtime error (MEMORY_DAEMON_UNAVAILABLE, MEMORY_DB_CORRUPT, MEMORY_RECOVERY_REQUIRED), record that incident explicitly instead of acting like retrieval returned an empty result.
• Treat degraded memory as blocked retrieval evidence for non-trivial work.
• Escalate to the adaptive memory runbook at ADAPTIVE_MEMORY_RUNBOOK.md when the runtime is degraded.
• Do not save new memory entries until the daemon and DB health are restored.

Trigger Reliability Harness

Use this skill-level eval corpus and keep thresholds fail-closed:

• false_positive_rate <= 0.10
• false_negative_rate <= 0.10

Positive Trigger Examples (should_trigger=true)

• "Learn how I code and apply my preferred testing/style choices automatically."
• "Use my saved frontend iteration preferences and update the implementation plan."
• "Resolve project preferences before coding this multi-file refactor."

Negative Trigger Examples (should_trigger=false)

• "What time is it?"
• "Summarize this paragraph."
• "Translate this sentence to Spanish."

If rates exceed thresholds, block approval and fix boundaries/corpus before rollout.