migrate-memory

Memory Organization & Migration

Maintain the auto-memory space: prune stale files, condense MEMORY.md index, and optionally migrate content to khive persistent memory.

When to Use

• MEMORY.md exceeds 150 lines (200-line truncation limit)
• Memory files accumulate without pruning
• User asks to organize, clean, or consolidate memory
• Switching to khive recall-based architecture (optional)

Architecture

MEMORY.md (index, <200 lines) ← loaded every session
  └→ *.md files (detail) ← loaded on demand via links
  └→ khive memory (semantic search) ← recalled via memory.recall()

All three layers coexist. MEMORY.md is the routing table — it tells Leo what exists and where to find detail. Individual files hold the content. khive memory is for cross-session semantic search.

What stays in MEMORY.md (always loaded)

• Recovery protocol (how to bootstrap)
• Health gates (P0 — blocking)
• Core directive
• People table (key relationships + rules)
• Critical life items (immigration, financial, commitments)
• Alignment table + behavioral correction links
• Active work table
• Recall triggers (query templates)

What goes in individual files

• Detailed project state (receipt numbers, specs, architecture)
• Feedback with context (why + how to apply)
• Reference data (API keys, restaurant lists, device IDs)

What goes in khive memory (optional)

• Cross-session learnings (KEY_INSIGHT)
• Session summaries (episodic)
• Patterns discovered across projects (semantic)

Organization Protocol

Phase 1: Inventory

MEMORY_DIR="$(find ~/.claude/projects -path '*/memory/MEMORY.md' -exec dirname {} \;)"
ls -lhS "$MEMORY_DIR"/*.md
wc -l "$MEMORY_DIR/MEMORY.md"

Phase 2: Audit

For each file (excluding MEMORY.md):

1. Read frontmatter (name, type, description)
2. Assess: ACTIVE | STALE | CONSOLIDATE
3. Check for dangling references in MEMORY.md

Use an Explore agent for bulk reading if >20 files.

Phase 3: Act

Stale files: Delete + remove from MEMORY.md index. Consolidate: Merge related files → update MEMORY.md link. Dangling refs: Create missing files or remove broken links. Bloated MEMORY.md: Move detail into individual files, keep 1-line pointer. Outdated info: Update dates, day counts, completed milestones.

Phase 4: Condense MEMORY.md

Target: <180 lines (safe margin under 200).

Techniques:

• Move detailed tracking (receipt numbers, timelines) to files
• Group behavioral corrections as compact link lists
• Condense tables (drop redundant columns)
• Remove past milestones that are complete
• Update stale dates to current values

Phase 5: Verify

• wc -l MEMORY.md < 200
• All linked files exist
• No dangling references
• Key info (health, people, commitments) still present

Phase 6: Optional khive migration

If khive memory is available and desired:

mcp__khive__memory(
  action="remember",
  content="[name]: [content]",
  memory_type="semantic|episodic",
  importance=0.70-0.85,
  source="auto-memory-migration-YYYYMMDD"
)

Tag with consistent source for rollback via forget_batch.

Quality Gates

• All files read before any deletion
• MEMORY.md stays under 200 lines
• No orphaned files (every .md linked or self-contained)
• Key sections preserved: health, people, alignment, active work