memory-retrievelisted

# Memory Retrieve ## Purpose / When to Activate Activate before `context-building` whenever a task requires historical context. **Never load full memory. Always filter by relevance.** **Never substitute summaries for durable memory (decisions, patterns, project).** --- ## 3-Level Progressive Disclosure ``` Level 1 — INDEX SCAN (~5 tokens/entry) Read index.md → Decision Registry table (DEC | Domain | Level | Title) Agent identifies relevant decision(s) by domain and/or level Level 2 — INDIVIDUAL ADR FILES (~300 tokens/file) Load specific decisions/DEC-{ID}.md files for full entry text Load other relevant category files (patterns, project, ops) Optionally traverse `related_to` in loaded files to discover adjacent decisions Or invoke `memory-search` Mode C for systematic cross-reference discovery Level 3 — CROSS-DOMAIN SCAN (only for Decision Consistency Gate) Grep frontmatter across all DEC-*.md files for conflicts Only triggered when recording a new decision (see Decision Consistency Gate in decision-extraction skill) ``` --- ## Process 1. **Read memory index** (`contexts/memory/index.md`). This contains: - Shared categories table (paths + purpose) - Decision Registry: one row per DEC-ID with domain, level, and title If `index.md` is absent or empty, fall back to scanning `contexts/memory/` directory structure. 2. **Identify relevant decisions** for the current task: - Filter the Decision Registry by **domain** (e.g., `billing`, `matchin