| name | permanent-memory-loader |
| description | Loads permanent_memory.md summary lines (≤120 always-loaded tokens, hard cap 200). Surfaces durable wisdom from past Epics into the current session. Silent if .claude/memory/ does not exist. |
| when_to_use | Every Claude turn when .claude/memory/permanent_memory.md exists; loads section headers + 1-line summaries only, never the full body |
| autoInvoke | true |
| allowed-tools | Read, Glob |
| effort | low |
| user-invocable | false |
AI-consumed reference. Optimized for Claude to read during execution.
Human-readable explanation: see docs/architecture/HIERARCHICAL_PLANNING.md
or 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)
- Detect: if
.claude/memory/permanent_memory.md does NOT exist → exit silently (no overhead)
- 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)
- Compose a compact context block: 1 line per Epic, 1 sub-line per subsection (≤120 chars each)
- Stamp with
trust: file per memory-trust-policy.md (memory content is read-verified, not user-canonical)
Token budget
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:
- Drop Tentative subsection lines first
- Trim "Patterns that worked" lines (keeping decisions + gotchas)
- Show Epic IDs only — no subsection lines
- Skip entirely if file > 8000 tokens (extreme; suggests epic-summarizer cap was bypassed)
What this skill loads (example output)
[permanent-memory | trust:file]
Epic FEAT-007 (JWT auth):
decisions: token storage in httpOnly cookie (DEC-001)
gotchas: refresh-token rotation requires CSRF re-issue
Epic FEAT-008 (Profile avatars):
decisions: deferred to S3 over base64 inline
anti-patterns: never store image bytes in postgres TEXT
What this skill does NOT do
- Does NOT load the full Epic body — only summary lines
- Does NOT load
.claude/memory/manual_overrides.md (that's user-curated; treated separately)
- Does NOT modify memory files (read-only)
- Does NOT make decisions based on memory content (caller decides)
- Does NOT trigger across projects (project-isolated; per
.claude/memory/ is per-project)
Memory staleness
If an Epic section is older than 90 days (parsed from its **Completed:** line), the loader prefixes it with (>90d may be outdated). This is per spec §29 risk register: "Permanent memory stale wisdom".
Tie-Ins
- Spec: §9.6, §19.2 (permanent_memory.md structure)
- Producer:
agents/epic-summarizer.md — only writer of the content this skill loads
- Producer (manual):
.claude/memory/manual_overrides.md — user-curated; loaded separately by name (not via this skill)
- Rule:
rules/core/memory-trust-policy.md — defines trust: file semantics for loaded content
- Rule:
rules/workflow/session-reset-policy.md — defines what's preserved across reset (this content is)
- Companion auto-invoke skill:
plan-loader — both fire silently when their files don't exist
