memory curator hook + /memory-audit (0.12.0)

[illinigirl]

The auto-memory twin of the CLAUDE.md curator — the 5th nudge-hook + judgment-skill pair, suggested from a review of the official memory docs.

Why

Auto memory's MEMORY.md (in ~/.claude/projects/<project>/memory/) loads only its first 200 lines / 25 KB each session — the rest is silently not loaded. A hard truncation, not a soft budget (the same shape as a scan dropping rows past page 1). And Claude writes this file itself across sessions, so it grows on its own, leaving tail entries that look stored but never reach context.

What's here

• memory-curator/ hook (PostToolUse) — fires only on MEMORY.md edits inside a memory dir (line + byte count, no LLM). Over MEMORY_LINE_BUDGET (200) or MEMORY_BYTE_BUDGET (25 KB) → warns the tail is silently dropped, nudges /memory-audit. Debounced; MEMORY_AUDIT_DISABLED off-switch; never auto-edits; fail-open. Topic-file edits and a bare repo MEMORY.md (not in a memory/ dir) don't fire.
• /memory-audit skill — reports what's below the cutoff (dropped now), staleness, redundancy/contradiction, and the index↔topic split; on approval tightens MEMORY.md to a lean index and moves detail to on-demand topic files (same index/chapters principle as CLAUDE.md).
• New memory-index-overflow catalog entry (hard-cutoff sibling of context-doc-bloat).
• Wired into hooks.json + CI; README section + layout; plugin & marketplace → 0.12.0; memory-audit-DESIGN.md.

Verification

• memory-curator 12/12 (line + byte cutoff, debounce, memory-dir detection incl. no false-fire on a repo MEMORY.md or topic files, disable, fail-safes).
• All 6 hook suites green; ruff clean; manifests valid; claude plugin validate passes.

🤖 Generated with Claude Code