Memory Management

Memory Architecture (Claude Code)

~/.claude/projects/<project>/memory/
  MEMORY.md              # Index file — pointers to memory files
  user_*.md              # User profile memories
  feedback_*.md          # Working style / correction memories
  project_*.md           # Active project context
  reference_*.md         # External system pointers

Each memory file has frontmatter:

---
name: descriptive name
description: one-line summary for relevance matching
type: user|feedback|project|reference
---

MEMORY.md is the index — one-line pointers, not content. Keep under 200 lines.

Operations

1. Daily Session Hygiene

At session end (or via obsidian-memory-sync skill):

• Update relevant memory files with new learnings
• Create new memory files for genuinely new topics
• Update MEMORY.md index if files were added/removed
• Never duplicate — check existing memories before creating new ones

2. Weekly Rollup (Every 7 Days)

Review all memory files:

• Merge related memories that have grown fragmented
• Update project memories with current status
• Verify reference memories still point to valid resources
• Remove memories that are now captured in code/git history

3. Monthly Archive

1. Review all project memories — archive completed projects
2. Check feedback memories — are they still relevant?
3. Verify user memories — has anything changed?
4. Audit MEMORY.md index — remove dead links, ensure all files are indexed

4. Cross-Reference Protocol

When writing new memories, check for connections:

• Does this relate to an existing project? -> Update that memory instead
• Does this contradict something previously noted? -> Flag and resolve
• Does this complete/update an open question? -> Close the loop
• Is this a pattern we've seen before? -> Note the pattern

5. Pruning Rules

Keep forever:

• User preferences and working agreements
• Business decisions and rationale
• Important relationships and context
• Lessons learned from mistakes
• Financial data and commitments

Prune after 30 days:

• Routine task completion notes
• Transient debugging sessions
• Superseded information (old plans replaced by new ones)

Prune after 90 days:

• Resolved blockers
• Completed one-off tasks
• Temporary workarounds that were properly fixed

Never store (use code/git instead):

• Code patterns, conventions, architecture, file paths
• Git history, recent changes, who-changed-what
• Debugging solutions or fix recipes
• Anything already in CLAUDE.md files

6. MEMORY.md Index Curation

Keep the index scannable:

• Group by type (User, Feedback, Projects, References)
• Each entry: - [Title](file.md) — one-line hook (under 150 chars)
• Alphabetize within groups
• Total lines under 200

7. Search & Retrieval

When looking for past context:

1. Check MEMORY.md index for relevant pointers
2. Read the specific memory file
3. If not found, grep across all memory files
4. Only if still not found, check Obsidian vault for partner agent memories

Maintenance Schedule

Integration Points

• self-improvement: Weekly review triggers memory rollup
• know-me: User preferences detected get stored via memory management protocols
• obsidian-memory-sync: Session end captures cross-skill context to Obsidian vault
• All skills: Memory captures cross-skill context (VPS sessions, store builds, research)