Claude-Mind: Skill for Claude Code

/aspirations-curate-memory — Daily Curation Sweep

Walks the bottom of the guardrail and reasoning-bank populations, retiring entries that have accumulated no positive evidence over many exposures and time, keeping those with recent signal, and routing genuinely-broken triggers to a Maintain goal.

This is the procedure that g-115-348 (daily, recurring, interval 24h) runs. The agent can also invoke it directly when the curation-log shows a backlog.

Why this exists

Audit on 2026-05-09 found 320 active guardrails and 693 active reasoning-bank entries; 143 of the guards and 378 of the rb entries had zero positive evidence (no helpful, no inferred, no active, no cited) across hundreds of LLM exposures. No retirement protocol fired. The framework was additive without being subtractive. This skill is the subtractive arm.

Cadence

g-115-348 runs every 24h (interval_hours: 24). At 15 items/day per kind, the guardrail population walks once in ~3 weeks; rb in ~7 weeks. After the first sweep, the next_review_eligible_at exemption window self-throttles daily volume to whatever has aged out (14d for zero-evidence, 30d for positive).

Procedure

Phase 0 — Load conventions

Bash: load-conventions.sh reasoning-guardrails learning-routing (no-op if already loaded).

Phase 1 — Pull candidates

Bash: bash core/scripts/utilization-stats.sh guardrails candidates --limit 15

Parse JSON. Each candidate item carries:

• id, evidence (composite score, see utilization-stats.py)
• retrieval_count, times_helpful, times_inferred_helpful, times_active, times_cited, times_skipped, times_inferred_unknown
• age_days, created
• next_review_eligible_at, auto_flagged_for_review
• category, rule (truncated to 140 chars)

If candidate_count == 0: skip to Phase 4. Routine close — population is caught up.

Phase 2 — Per-item decision

For each candidate (process in returned order — already sorted by evidence asc, age desc, exposure desc):