permanent-memory-loader

Permanent Memory Loader

Auto-invoke complement to epic-summarizer. Read-only.

Behavior (in order)

1. Detect: if .claude/memory/permanent_memory.md does NOT exist → exit silently
2. Read the file. For each ## Epic: <ID> — <intent> section, apply the dual-read rule (below) to load only its heading + one summary line — never the full body / subsections.
3. Compose a compact block: 1 line per Epic (heading + its one summary line, ≤120 chars)
4. Stamp trust: file per rules/core/memory-trust-policy.md

Dual-read rule (v2 fast-path + v1 fallback)

Sections come in two shapes; both load, and the loader picks per section:

• v2 (fast-path): if the section body's first line is an <!-- af-memory:v2 epic=… confidence=… items=… --> marker, load exactly three lines — the ## heading, that marker line, and the single one-line human summary immediately after it. Never read the prose subsections or typed entries.
• v1 (fallback, backward-compat): if the section has no af-memory:v2 marker, load the ## heading + the first non-empty line after it. Existing v1 files are never rewritten or migrated — they keep working unchanged.

Exact lines loaded per section:

dualread[2]{shape,detect,lines_loaded}:
  v2,"body starts with <!-- af-memory:v2 ... -->","## heading + marker line + next 1-line summary"
  v1,"no af-memory:v2 marker","## heading + first non-empty line"

The marker's confidence=/items= fields are available cheaply for degradation decisions without touching the body.

Token budget

budget[3]{layer,target,hard_cap}:
  always_loaded,120,200
  per_epic_summary,12,20
  per_subsection_line,1,2

Budget exceeded → degrade (drop Tentative → trim patterns → Epic IDs only → skip if file > 8000 tokens).

Constraints

Never loads full Epic bodies or manual_overrides.md; never writes; never decides from memory content. Epics older than 90 days are prefixed (>90d may be outdated).