Skill Metadata

What It Does

Session memory 및 context 관리 전략을 제공합니다. Just-in-time retrieval, insight caching, memory file 관리를 통해 context window를 효율적으로 사용하는 방법을 다룹니다.

When to Use

• Context limit에 도달할 위험이 있을 때
• 대규모 프로젝트에서 효율적인 context 관리가 필요할 때
• Session handoff를 준비할 때
• Memory file 구조를 설계하거나 정리할 때

Managing Claude Code Session Memory & Context

Claude Code operates within context windows (~100K-200K tokens). Effective memory management ensures productive sessions without hitting limits.

Context Budget Overview

Total Context Budget
├── System Prompt (~2K)
├── Tools & Instructions (~5K)
├── Session History (~30K)
├── Project Context (~40K)
└── Available for Response (~23K)

Just-in-Time (JIT) Retrieval Strategy

High-Freedom: Core Principles

Principle 1: Pull Only What You Need

• Don't load entire codebase upfront
• Load files relevant to immediate task
• Use Glob/Grep for targeted searches
• Cache results for reuse

Principle 2: Prefer Explore Over Manual Hunting

# ❌ Manual approach: Search many files, load all
rg "authenticate" src/ | head -20

# ✅ JIT approach: Use Explore agent
@agent-Explore "Find authentication implementation, analyze"

Principle 3: Layered Context Summaries

1. High-level brief (purpose, success criteria)
   ↓
2. Technical core (entry points, domain models)
   ↓
3. Edge cases (known bugs, constraints)

Example: Feature Implementation

Task: "Add email verification to signup"

JIT Retrieval:
├── Read: User model (src/domain/user.ts)
├── Read: Signup endpoint (src/api/auth.ts)
├── Grep: "email" in tests (understand patterns)
├── Glob: Find email service (src/infra/email.*)
└── Cache: Signup flow diagram in memory

Medium-Freedom: Memory File Patterns

Pattern 1: Session Summary Cache

File: .moai/memory/session-summary.md

# Session Summary

## Current Task
- Feature: User email verification
- SPEC: AUTH-015
- Status: In RED phase (writing tests)

## Key Files
- Test: tests/auth/email_verify.test.ts
- Impl: src/domain/email_service.ts
- Config: src/config/email.ts

## Important Context
- Email service uses SendGrid API
- Verification tokens expire in 24h
- Already have similar flow for password reset (AUTH-012)

## Assumptions Made
- Assuming transactional emails only
- Async email sending OK
- No SMS verification needed

Pattern 2: Architecture Reference

File: .moai/memory/architecture.md

# Architecture Reference

## Data Flow for Email Verification

User(Browser) ↓ [POST /auth/signup] Server ↓ [Create user + token] DB ↓ [sendEmail async] Queue ↓ [Process job] Email Service (SendGrid) ↓ User receives email with link User clicks link ↓ [GET /auth/verify?token=...] Server validates token ↓ [Mark user verified] DB ↓ User logged in

## Module Boundaries
- `domain/`: Business logic (no framework)
- `api/`: HTTP endpoints only
- `infra/`: External services (SendGrid, DB)

Pattern 3: Known Gotchas Cache

File: .moai/memory/gotchas.md

# Common Pitfalls in This Project

## Email Service
- SendGrid has rate limit: 100 emails/sec per account
- Test mode uses fake email (won't actually send)
- Async job failures don't alert (check logs)

## Database
- Migrations must be reviewed before prod deploy
- Test DB is reset after each suite
- Foreign key constraints enforced (plan deletions)

## Authentication
- JWT tokens stored in httpOnly cookies (XSRF protected)
- Refresh token rotation required (not automatic)
- Session timeout: 7 days (hardcoded, not configurable yet)

Low-Freedom: Memory Management Practices

Practice 1: Caching Key Insights

After reading code:
1. Note file locations (~5 min read)
2. Summarize key logic (~2 min)
3. Write to memory file (~1 min)
4. Reference in next session

Example memory entry:

# USER-002: Email verification flow

## Key Code Locations
- Token generation: src/domain/user.ts:generateVerificationToken()
- Email sending: src/infra/email_service.ts:sendVerificationEmail()
- Token validation: src/api/auth.ts:POST /verify

## Logic Summary
1. User submits email → server generates token (16 chars, base64)
2. Token stored in DB with 24h expiry
3. Email sent async via SendGrid
4. User clicks link → token validated → user marked verified
5. Token deleted after use (can't reuse)

## Related TESTs
- tests/auth/email_verify.test.ts (GREEN phase - needs implementation)
- Similar flow: password reset (PASSWORD-001)

Practice 2: Session Boundary Management

Before switching between tasks:

# Session Handoff Note

## Completed
✓ RED phase: 3 test cases for email verification
✓ GREEN phase: Minimal implementation passing tests
✓ REFACTOR: Added input validation

## Status
- Current: Ready for /alfred:3-sync
- Next action: Run full test suite, then sync docs

## Context for Next Session
- SPEC: .moai/specs/SPEC-AUTH-015/spec.md
- Tests: tests/auth/email_verify.test.ts (all passing)
- Code: src/domain/email_service.py
- Database migration: pending (see migrations/ directory)

## Assumptions
- SendGrid API key set in .env
- Test mode uses mock email service
- Database schema includes email_verified_at column

Practice 3: Cleanup Before Session End

# Remove unnecessary cached files
rm .moai/memory/temp-*.md

# Archive completed memory files
mv .moai/memory/feature-x-* .moai/memory/archive/

# Keep only active session memory
ls -la .moai/memory/
# session-summary.md (current)
# architecture.md (reference)
# gotchas.md (patterns)

Memory File Organization

.moai/
├── memory/
│   ├── session-summary.md      # Current session state
│   ├── architecture.md         # System design reference
│   ├── gotchas.md             # Common pitfalls
│   ├── spec-index.md          # List of all SPECs + status
│   ├── api-reference.md       # API endpoints quick lookup
│   └── archive/               # Completed session notes
│       ├── feature-auth-*
│       └── feature-api-*
└── specs/                      # Requirement specifications
    ├── SPEC-AUTH-001/
    ├── SPEC-USER-002/
    └── SPEC-API-003/

Context Optimization Checklist

• Memory files describe architecture (not code)
• Session summary updated before handoff
• Key file locations cached (don't re-search)
• Assumptions explicitly documented
• No duplicate information between memory files
• Archive files moved after session completion
• All cached insights reference file paths
• Memory files are Markdown (human-readable)

Best Practices

✅ DO:

• Use Explore for large searches
• Cache results in memory files
• Keep memory files < 500 lines each
• Update session-summary.md before switching tasks
• Reference memory files in handoff notes

❌ DON'T:

• Load entire src/ or docs/ directory upfront
• Duplicate context between memory files
• Store memory files outside .moai/memory/
• Leave stale session notes (archive or delete)
• Cache raw code (summarize logic instead)

Commands for Memory Management

# View current session memory
cat .moai/memory/session-summary.md

# List all memory files
ls -la .moai/memory/

# Archive completed work
mv .moai/memory/feature-old-* .moai/memory/archive/

# Search memory files
grep -r "email verification" .moai/memory/

# Count context usage estimate
wc -w .moai/memory/*.md  # Total words

Reference: Claude Code Context Management Version: 1.0.0