Manage Memory

Maintain Claude Code's persistent memory dir → accurate, concise, useful across sessions. MEMORY.md loaded into system prompt every conv — lines after 200 truncated → file must be lean index pointing to topic files for detail.

Use When

• MEMORY.md nearing 200-line limit
• Session produced durable insights worth preserving (new patterns, arch decisions, debugging solutions)
• Topic section in MEMORY.md > 10-15 lines → extract
• Project state changed (renamed files, new domains, updated counts) → entries may be stale
• Starting new work area → check if relevant memory exists
• Periodic maintenance between sessions

In

• Req: Access to memory dir (typically ~/.claude/projects/<project-path>/memory/)
• Opt: Specific trigger ("MEMORY.md too long", "just finished major refactor")
• Opt: Topic to add / update / extract

Do

Step 1: Assess Current State

Read MEMORY.md + list all files in memory dir:

wc -l <memory-dir>/MEMORY.md
ls -la <memory-dir>/

Check line count vs 200-line limit. Inventory existing topic files.

→ Clear picture of total lines, # topic files, which sections exist in MEMORY.md.

If err: Memory dir doesn't exist → create. MEMORY.md doesn't exist → minimal one w/ # Project Memory header + ## Topic Files section.

Step 2: ID Stale Entries

Compare memory claims vs current project state. Common staleness:

1. Count drift: File counts, skill counts, domain counts changed after additions/removals
2. Renamed paths: Files / dirs moved / renamed
3. Superseded patterns: Workarounds no longer needed after fixes
4. Contradictions: Two entries saying diff things about same topic

Use Grep to spot-check:

# Example: verify a skill count claim
grep -c "^      - id:" skills/_registry.yml
# Example: verify a file still exists
ls path/claimed/in/memory.md

→ List of stale entries w/ correct current vals.

If err: Can't verify claim (refs external state) → leave but add (unverified) note rather than silently preserve potentially wrong info.

Step 3: Decide What to Add

For new entries, apply filters before writing:

1. Durability: Will this be true next session? Avoid session-specific (current task, in-progress, temporary).
2. Non-duplication: CLAUDE.md / project docs already cover? Don't duplicate — memory for things NOT captured elsewhere.
3. Verified: Confirmed across multi interactions, or single obs? Single → verify vs project docs before writing.
4. Actionable: Does knowing change behavior? "Sky blue" ≠ useful. "Exit code 5 = quoting err → use temp files" changes how you work.

Exception: User explicitly asks to remember → save immediately, no multi confirmations needed.

→ Filtered list worth adding, each meeting durability + non-dup + verified + actionable.

If err: Unsure if worth keeping → err toward keeping briefly in MEMORY.md — easier to prune later than rediscover.

Step 4: Extract Oversize Topics

Section > ~10-15 lines → extract to dedicated topic file:

1. Create <memory-dir>/<topic-name>.md w/ descriptive header
2. Move detailed content from MEMORY.md → topic file
3. Replace section in MEMORY.md w/ 1-2 line summary + link:

## Topic Files
- [topic-name.md](topic-name.md) — Brief description of contents

Naming conventions:

• Lowercase kebab-case: viz-architecture.md, not VizArchitecture.md
• Name by topic, not chronology: patterns.md, not session-2024-12.md
• Group related: combine "R debugging" + "WSL quirks" → patterns.md vs one file per fact

→ MEMORY.md stays < 200 lines. Each topic file self-contained + readable w/o MEMORY.md ctx.

If err: Topic file < 5 lines → probably not worth extracting → leave inline.

Step 5: Update MEMORY.md

Apply changes: remove stale, add new, update counts, ensure Topic Files section lists all dedicated files.

MEMORY.md structure:

# Project Memory

## Section 1 — High-level context
- Bullet points, concise

## Section 2 — Another topic
- Key facts only

## Topic Files
- [file.md](file.md) — What it covers

Guidelines:

• Each bullet 1-2 lines max
• Inline formatting (code, bold) for scanability
• Most frequently needed ctx first
• Topic Files section always last

→ MEMORY.md < 200 lines, accurate, working links to all topic files.

If err: Can't get < 200 after extraction → ID least-freq-used section + extract. Every section candidate — even project structure overview can go to topic file if needed, leaving 1-line summary.

Step 6: Verify Integrity

Final check:

1. Line count: MEMORY.md < 200
2. Links: Every topic file referenced exists
3. Orphans: Topic files not referenced in MEMORY.md
4. Accuracy: Spot-check 2-3 factual claims vs project state

wc -l <memory-dir>/MEMORY.md
# Check for broken links
for f in $(grep -oP '\[.*?\]\(\K[^)]+' <memory-dir>/MEMORY.md); do
  ls <memory-dir>/$f 2>/dev/null || echo "BROKEN: $f"
done
# Check for orphan files
ls <memory-dir>/*.md | grep -v MEMORY.md

→ Line count < 200, no broken links, no orphans, spot-checked claims accurate.

If err: Fix broken links (update / remove). Orphans → add ref in MEMORY.md / delete if no longer relevant.

Check

• MEMORY.md < 200 lines
• All referenced topic files exist on disk
• No orphan .md in memory dir (every file linked from MEMORY.md)
• No stale counts / renamed paths
• New entries meet durability / non-dup / verified / actionable
• Topic files have descriptive headers + self-contained
• MEMORY.md reads as quick-ref, not changelog

Traps

• Memory file pollution: Writing every session obs to memory. Most findings session-specific + don't need persisting. Apply 4 filters (Step 3) before writing.
• Stale counts: Updating code but not memory. Counts (skills, agents, domains, files) drift silently. Always verify vs source of truth before trusting memory.
• Chronological organization: "When I learned" vs "what it's about". Topic-based (patterns.md, viz-architecture.md) > date-based files.
• Duplicate CLAUDE.md: CLAUDE.md = authoritative project instructions. Memory captures things NOT in CLAUDE.md — debugging insights, arch decisions, workflow prefs, cross-project patterns.
• Over-extraction: Topic file for every 3-line section. Only extract when > ~10-15 lines. Small sections inline.
• Forget 200-line limit: MEMORY.md loaded every system prompt. Lines after 200 silently truncated. Grows past → bottom content effectively invisible.

→

• write-claude-md — CLAUDE.md captures project instructions; memory captures cross-session learning
• prune-agent-memory — inverse of manage-memory: auditing, classifying, selectively forgetting stored
• write-continue-here — structured continuation file for session handoff; complements memory as short-term bridge
• read-continue-here — read + act on continuation at session start; consumption side of handoff
• create-skill — new skills may produce memory-worthy patterns
• heal — self-healing may update memory as part of integration step
• meditate — meditation sessions may surface insights worth persisting