Memory patterns

Memory Patterns

Claude Code's auto memory system persists knowledge across sessions via files in the memory directory. MEMORY.md is always loaded into context; topic files store detailed notes referenced from MEMORY.md.

When to Use

• User asks "remember this" or "always do X"
• User corrects a mistake that came from memory — update the source
• You discover stable project patterns worth persisting
• User asks about organizing cross-session knowledge
• You need to structure MEMORY.md or create topic files

Core Patterns

MEMORY.md Structure

MEMORY.md is the index file — always loaded, so it must stay concise (under 200 lines). Use it as a table of contents pointing to detailed topic files:

# Project Memory

## Project Overview
- Name: acme-api
- Stack: Node.js, Express, PostgreSQL, TypeScript
- Test framework: Vitest
- Package manager: pnpm

## Key Conventions
- Use Zod for all input validation
- Repository pattern for database access
- All API responses follow ApiResponse<T> interface
- Feature branches: feat/description, fix/description

## Architecture
- See [architecture.md](./architecture.md) for service layout
- See [database.md](./database.md) for schema notes

## User Preferences
- Prefers functional style over classes
- Always use named exports (no default exports)
- Run tests before committing

## Known Issues
- See [debugging.md](./debugging.md) for recurring problems

Topic File Organization

Create separate files for detailed knowledge. Link them from MEMORY.md:

memory/
  MEMORY.md              # Index (always loaded, < 200 lines)
  architecture.md        # Service layout, module boundaries
  database.md            # Schema notes, migration patterns
  debugging.md           # Recurring issues and solutions
  patterns.md            # Code patterns specific to this project
  api-conventions.md     # API design decisions

Each topic file should be self-contained:

# Architecture Notes

## Service Layout
- src/services/ — Business logic, one file per domain
- src/routes/ — Express route handlers, thin layer
- src/db/ — Database access via repository pattern
- src/middleware/ — Auth, validation, error handling

## Module Boundaries
- Services never import from routes
- Routes call services, never DB directly
- Middleware is stateless

## Key Dependencies
- src/services/auth.ts depends on src/db/users.ts
- src/services/billing.ts depends on external Stripe SDK

When to Save

Save to memory when:

• A pattern is confirmed across multiple interactions (not just one file)
• The user explicitly asks to remember something
• An architectural decision affects future work
• A debugging solution applies to a recurring problem
• User preferences are stated ("always use X", "never do Y")

When NOT to Save

Do not save:

• Session-specific context (current task details, in-progress work)
• Information from a single file read that might be incomplete
• Speculative conclusions not verified against project docs
• Anything that duplicates CLAUDE.md instructions
• Temporary state or one-off fixes

Update vs Create

Before creating a new memory entry, check if an existing one covers the topic:

# WRONG: Creating duplicates
## Database Notes (added 2024-01-15)
Uses PostgreSQL with Prisma ORM.

## Database Info (added 2024-02-03)
Database is PostgreSQL. Uses Prisma for ORM.

# RIGHT: Update the existing entry
## Database
- PostgreSQL with Prisma ORM
- Migrations in prisma/migrations/
- Seed data in prisma/seed.ts

Handling Corrections

When the user corrects something you stated from memory, update the source immediately:

User: "No, we switched from Prisma to Drizzle last month"

# IMMEDIATELY update MEMORY.md or the relevant topic file:
# Old: "Database: PostgreSQL with Prisma ORM"
# New: "Database: PostgreSQL with Drizzle ORM"

Never continue with incorrect memory. Fix the source first, then proceed.

Semantic Organization

Group memories by topic, not by time:

# WRONG: Chronological
## January 15
Learned that project uses TypeScript.
## January 16
Found out tests use Vitest.

# RIGHT: Semantic
## Tech Stack
- Language: TypeScript
- Test framework: Vitest

Anti-Patterns

• Bloated MEMORY.md: Exceeding 200 lines means important information gets truncated. Move details to topic files and keep MEMORY.md as an index.
• Saving speculative information: Reading one file and concluding "the project uses pattern X" without verification leads to incorrect memories. Verify against multiple sources.
• Duplicating CLAUDE.md: If project instructions already exist in CLAUDE.md, do not repeat them in memory. Memory is for learned knowledge, not static configuration.
• Never updating: Memories rot. When you discover information is outdated, update or remove it immediately.
• Saving session state: "Currently working on the auth refactor" does not belong in memory. It is true now but meaningless next session.

Quick Reference

File	Purpose	Size Limit
MEMORY.md	Index, always loaded	< 200 lines
Topic files	Detailed knowledge	Unlimited

Save: Stable patterns, user preferences, architecture decisions, recurring solutions Skip: Session state, unverified guesses, CLAUDE.md duplicates Update: When corrected, when information changes, when patterns evolve Organize: By topic (semantic), not by date (chronological)