What problem does it solve? Long-running AI sessions accumulate valuable user preferences, environment details, and workflow patterns, but writing everything to memory causes bloat while writing nothing loses hard-won context. This Skill filters session transcripts through a three-question test (growth direction, framework dimension, fermentation potential) so only durable, high-signal facts are persisted within a strict 2,200-character budget. ## Core Features & Use Cases - Three-Question Filtering: Each candidate memory must answer whether it supports system growth, maps to a framework dimension, and has fermentation potential before being saved. - Constitutional Guardrails: Memories that conflict with immutable CONSTITUTION clauses are rejected, and imperative phrasing ("Always use X") is rewritten as declarative facts ("User prefers X"). - Atomic Capacity Management: Updates execute in remove→replace→add order with all-or-nothing semantics, and usage above 80% triggers proactive cleanup to land in the 60-75% target range. - Use Case: After a 12-tool-call session that completes a literature deduplication pipeline, the Skill removes one outdated path entry, adds two annotated entries about the new path and sorting preference, and reports memory usage dropping from 82% to 71%. ## Quick Start Review this session transcript and my current memory, then save only the entries that pass the three-question filter while keeping usage under 75%.
