Skip to content

Claude Skills - Memory Management

One-Liner

Memory Management Skills is a cross-session knowledge persistence system — Through Memory compression, Tips recording, and Memory updates, AI remembers project context across sessions.

Pain Points Solved

Pain PointCurrent StateClaude Code Workflow Solution
New session amnesiaEvery conversation needs to re-explain project backgroundMemory persists context
Knowledge lossValuable insights and decisions disappear with sessionMemory compression and Tips recording
Context window limitsContext exceeds window after long conversationsMemory extraction and merging
Difficult knowledge retrievalHard to find historical recordsMemory search and embedding

Skills List

SkillFunctionTrigger
memory-captureUnified memory capture (session compression/quick notes)/memory-capture
memory-manageMemory update (full/related/single)/memory-manage

Skills Details

memory-capture

One-Liner: Unified memory capture — Dual-mode routing for session compression or quick notes

Trigger:

/memory-capture                              # Interactive mode selection
/memory-capture compact                      # Session compression mode
/memory-capture tip "Note content"           # Quick note mode
/memory-capture "Use Redis" --tag config     # Note with tags

Features:

  • Dual-mode routing: Auto-detects user intent, routes to compression mode or notes mode
  • Compact mode: Compresses complete session memory to structured text for session recovery
  • Tips mode: Quickly records ideas, snippets, insights

Architecture Overview:

┌─────────────────────────────────────────────┐
│  Memory Capture (Router)                    │
│  → Parse input → Detect mode → Route to phase│
└───────────────┬─────────────────────────────┘

        ┌───────┴───────┐
        ↓               ↓
  ┌───────────┐   ┌───────────┐
  │  Compact  │   │   Tips    │
  │  (Phase1) │   │  (Phase2) │
  │  Full     │   │  Quick    │
  │  Session  │   │  Note     │
  └─────┬─────┘   └─────┬─────┘
        │               │
        └───────┬───────┘

        ┌───────────────┐
        │  core_memory  │
        │   (import)    │
        └───────────────┘

Auto Routing Rules (priority order):

SignalRouteExample
Keywords: compact, session, 压缩, recovery→ Compact"Compress current session"
Keywords: tip, note, 记录, 快速→ Tips"Record an idea"
Has --tag or --context flag→ Tips"note content" --tag bug
Short text (<100 chars) + no session keywords→ Tips"Remember to use Redis"
Ambiguous or no parametersAskUserQuestion/memory-capture

Compact Mode:

  • Use case: Compress current complete session memory (for session recovery)
  • Input: Optional "session description" as supplementary context
  • Output: Structured text + Recovery ID
  • Example: /memory-capture compact or /memory-capture "Completed authentication module"

Tips Mode:

  • Use case: Quickly record a note/idea/tip
  • Input:
    • Required: <note content> - Note text
    • Optional: --tag <tag1,tag2> categories
    • Optional: --context <context> associated code/feature reference
  • Output: Confirmation + ID + tags
  • Example: /memory-capture tip "Use Redis for rate limiting" --tag config

Core Rules:

  1. Single-phase execution: Each call executes only one phase — never both
  2. Content fidelity: Phase files contain complete execution details — follow exactly
  3. Absolute paths: All file paths in output must be absolute paths
  4. No summarization: Compact mode preserves complete plan — never abbreviate
  5. Speed priority: Tips mode should be fast — minimal analysis overhead

memory-manage

One-Liner: Memory update — Full/related/single update modes

Trigger:

/memory-manage                               # Interactive mode
/memory-manage full                          # Full update
/memory-manage related <query>               # Related update
/memory-manage single <id> <content>         # Single update

Features:

  • Three update modes: Full update, related update, single update
  • Memory search and embedding
  • Memory merge and compression

Update Modes:

ModeTriggerDescription
fullfull or -fRegenerate all Memory
relatedrelated <query> or -r <query>Update Memory related to query
singlesingle <id> <content> or -s <id> <content>Update single Memory entry

Memory Storage:

  • Location: C:\Users\dyw\.claude\projects\D--ccw-doc2\memory\
  • File: MEMORY.md (main memory file, truncated after 200 lines)
  • Topic files: Independent memory files organized by topic

Memory Types:

TypeFormatDescription
CMEM-YYYYMMDD-HHMMSSTimestamp formatTimestamped persistent memory
Topic filesdebugging.md, patterns.mdMemory organized by topic

Core Rules:

  1. Prefer update: Update existing memory rather than writing duplicate content
  2. Topic organization: Create independent memory files categorized by topic
  3. Delete outdated: Delete memory entries that are proven wrong or outdated
  4. Session-specific: Don't save session-specific context (current task, in-progress work, temporary state)

Best Practices

  1. Session compression: Use memory-capture compact after long conversations to save complete context
  2. Quick notes: Use memory-capture tip to quickly record ideas and insights
  3. Tag categorization: Use --tag to add tags to notes for later retrieval
  4. Memory search: Use memory-manage related <query> to find related memories
  5. Regular merging: Regularly use memory-manage full to merge and compress memories

Memory File Structure

memory/
├── MEMORY.md                 # Main memory file (line limit)
├── debugging.md              # Debugging patterns and insights
├── patterns.md               # Code patterns and conventions
├── conventions.md            # Project conventions
└── [topic].md               # Other topic files

Usage Examples

bash
# Compress current session
/memory-capture compact

# Quickly record an idea
/memory-capture tip "Use Redis for rate limiting" --tag config

# Record note with context
/memory-capture "Authentication module uses JWT" --context "src/auth/"

# Update related memories
/memory-manage related "authentication"

# Full memory update
/memory-manage full

Released under the MIT License.