Claude Code Memory & Instructions System

Claude Code Memory & Instructions System

Complete guide to CLAUDE.md, project instructions, auto-memory, and rules.

CLAUDE.md — Project Instructions

Purpose

CLAUDE.md files contain natural-language instructions that Claude Code loads as system context. They tell Claude about your project's conventions, build commands, architecture, and preferences.

Hierarchy (All Additive)

Enterprise/Managed CLAUDE.md     (highest priority, cannot be overridden)
  └── User ~/.claude/CLAUDE.md         (personal preferences, all projects)
      └── Project ./CLAUDE.md              (project conventions)
          └── Project ./.claude/CLAUDE.md       (additional project instructions)
              └── Local ./CLAUDE.local.md         (personal, gitignored)
                  └── Directory path/CLAUDE.md      (scoped to that directory)

All levels are loaded and combined. Lower levels cannot override enterprise settings.

Additional Syntax

• @path/file.md — Import additional files into CLAUDE.md context
• Keep under 200 lines for optimal context cost

What to Include

# Project Instructions

## Build & Test Commands
- Install: `pnpm install`
- Dev: `pnpm dev`
- Build: `pnpm build`
- Test: `pnpm test`
- Single test: `pnpm test -- path/to/test.ts`
- Lint: `pnpm lint`
- Type check: `pnpm typecheck`

## Code Conventions
- Language: TypeScript strict mode
- Style: Prettier + ESLint
- Imports: Use absolute paths from `@/`
- Components: Functional React with hooks
- State: Zustand for global, React Query for server

## Architecture
- Monorepo: Turborepo with packages in `packages/`
- API: Next.js API routes in `src/app/api/`
- Database: PostgreSQL with Prisma ORM
- Auth: NextAuth.js with Keycloak provider
- Deployment: Kubernetes on AWS EKS

## Key Files
- `prisma/schema.prisma` — Database schema
- `src/lib/auth.ts` — Auth configuration
- `src/lib/db.ts` — Database client

## Important Context
- We use pnpm, never npm or yarn
- Always run tests before committing
- Follow conventional commits format
- Never commit .env files

What NOT to Include

• Secrets or API keys
• User-specific paths (use settings.local.json instead)
• Temporary instructions (use conversation context)
• Overly detailed documentation (link to docs instead)

Rules System

Rules in .claude/rules/ are like scoped CLAUDE.md files.

Global Rules (no frontmatter)

# Git Rules
- Use conventional commits
- Never force push to main

Path-Scoped Rules

---
paths:
  - "**/*.ts"
  - "**/*.tsx"
---

# TypeScript Rules
- Use strict mode
- Prefer interfaces over type aliases for objects
- Max function length: 50 lines

Rule File Organization

.claude/rules/
├── architecture.md      # Global — project structure
├── code-style.md        # Scoped — *.ts, *.tsx, *.js, *.jsx
├── docker-k8s.md        # Scoped — Dockerfiles, *.yaml
├── git-workflow.md      # Global — commit conventions
├── testing.md           # Scoped — *.test.*, *.spec.*
├── research.md          # Global — web research tools
├── self-healing.md      # Global — error capture protocol
└── lessons-learned.md   # Global — auto-growing error log

Auto-Memory System

How It Works

Claude Code automatically saves important discoveries and patterns to persistent memory files.

Memory Location

~/.claude/projects/<project-hash>/memory/
├── MEMORY.md        # Auto-loaded into context (keep under 200 lines)
├── debugging.md     # Topic-specific notes
├── patterns.md      # Recurring patterns
└── architecture.md  # Architecture decisions

MEMORY.md (Auto-Loaded)

# Project Memory

## Build System
- Uses pnpm workspaces with Turborepo
- `pnpm -w` for root-level operations

## Database
- PostgreSQL on port 5433 (non-standard)
- Run `pnpm db:migrate` after schema changes

## User Preferences
- Always use bun for running scripts
- Prefers functional style over OOP

What Gets Auto-Saved

• Stable patterns confirmed across interactions
• Key architectural decisions
• Important file paths
• User preferences for workflow
• Solutions to recurring problems

What Does NOT Get Saved

• Session-specific context
• Unverified information
• Anything duplicating CLAUDE.md
• Speculative conclusions

Manual Memory Commands

# User says:
"Remember that we always use bun, not npm"
→ Claude saves to MEMORY.md

"Forget the instruction about yarn"
→ Claude removes from MEMORY.md

"Always run lint before committing"
→ Claude saves preference

# Slash commands:
/memory              # View memories
/memory add <text>   # Add entry
/memory clear        # Clear all

Disabling Auto-Memory

export DISABLE_AUTOMEMORY=1

or in settings.json:

{ "autoMemory": false }

Self-Healing Loop

The lessons-learned system creates a continuously improving knowledge base:

Tool call fails
  → PostToolUse hook captures error
  → Appends to .claude/rules/lessons-learned.md
  → Claude fixes the issue
  → Claude updates entry: Status → RESOLVED, adds Fix + Prevention
  → Next session loads lessons-learned.md as a rule
  → Claude doesn't repeat the mistake

Lessons-Learned Format

### Error: Git push rejected
- **Tool:** Bash
- **Status:** RESOLVED
- **Fix:** Use `git pull --rebase` before pushing
- **Prevention:** Always pull before push

Settings Hierarchy

Enterprise managed settings     (cannot be overridden)
  └── User settings (~/.claude/settings.json)
      └── Project settings (.claude/settings.json)
          └── Local overrides (.claude/settings.local.json)

Merge Behavior

• Arrays: concatenated (allow lists combine)
• Objects: deep merged (later values override)
• Primitives: later values override
• Enterprise settings: always win

Best Practices

1. Keep CLAUDE.md focused — Build commands, conventions, architecture
2. Use rules for enforcement — Path-scoped rules for file-type conventions
3. Let auto-memory learn — Don't manually duplicate auto-saved info
4. Review lessons-learned — Promote patterns to permanent rules
5. Scope rules narrowly — Docker rules shouldn't load for TypeScript work
6. Version control CLAUDE.md — It's project documentation
7. Gitignore settings.local.json — Personal overrides stay local