permanent-memory-loader

> **AI-consumed reference.** Optimized for Claude to read during execution. > Human-readable explanation: see [docs/architecture/HIERARCHICAL_PLANNING.md](../../../docs/architecture/HIERARCHICAL_PLANNING.md) > or [docs/getting-started/](../../../docs/getting-started/) depending on topic. # Permanent Memory Loader **STATUS — v3.7.0-alpha.4 (Milestone C interim).** Auto-invoke complement to `epic-summarizer`. ## Behavior (in order) 1. **Detect:** if `.claude/memory/permanent_memory.md` does NOT exist → exit silently (no overhead) 2. **Read** the file, extract: - Each `## Epic: <ID> — <intent>` header - For each Epic, the first non-empty line of each subsection (Architectural decisions, Gotchas discovered, Anti-patterns to avoid, Patterns that worked, Conflicts encountered) 3. **Compose** a compact context block: 1 line per Epic, 1 sub-line per subsection (≤120 chars each) 4. **Stamp** with `trust: file` per `memory-trust-policy.md` (memory content is read-verified, not user-canonical) ## Token budget ```toon budget[3]{layer,target,hard_cap}: always_loaded,120,200 per_epic_summary,12,20 per_subsection_line,1,2 ``` Auto-degradation when budget exceeded: 1. **Drop Tentative subsection** lines first 2. **Trim "Patterns that worked"** lines (keeping decisions + gotchas) 3. **Show Epic IDs only** — no subsection lines 4. **Skip entirely** if file > 8000 tokens (extreme; suggests epic-summarizer cap was bypassed) ## What this skill loads (example output) ``` [per...