Claude Memory

Claude Memory

Overview

Claude Code has a built-in memory system that persists context across sessions. Without it, every session starts cold — Claude has no knowledge of past decisions, your preferences, or what was built last week. With a well-managed memory directory, Claude picks up exactly where you left off: knowing your role, your stack, your conventions, and the current state of the project.

When to Use

• Claude keeps forgetting important context between sessions
• Starting a new session and want to pick up exactly where you left off
• Building a system where Claude accumulates knowledge over time
• Managing what Claude should and shouldn't remember

Core Jobs

1. How Claude Code Memory Works

Claude Code has an auto-memory system in ~/.claude/projects/[project-hash]/memory/:

~/.claude/
  CLAUDE.md                    ← global instructions (always loaded)
  projects/
    [project-hash]/
      memory/
        MEMORY.md              ← index of all memory files (always loaded)
        user_profile.md        ← who the user is, preferences
        project_context.md     ← what the project is about
        feedback_*.md          ← corrections and learned preferences
        reference_*.md         ← external resources and links

Claude loads MEMORY.md on every session start and follows pointers to load relevant memory files. Keep the index concise — it is always loaded in full.

2. Memory Types

3. Writing Good Memory Entries Manually

---
name: user-coding-style
description: User's coding preferences and style requirements
type: user
---

User: Kien, Vietnamese communication preferred
Role: Solo AI builder, primarily Python + TypeScript
Experience: Senior, minimal explanation needed
Style: Direct, concise, no over-engineering
Key preference: Always use uv for Python package management, never pip
Test preference: pytest with fixtures, no mocking external services

Good memory entries are specific, actionable, and focused on preferences that change Claude's behavior. Avoid storing code snippets or file contents — those change and go stale.

4. Memory Index (MEMORY.md)

Keep the index concise — it is loaded every session and truncated if too long:

# Memory Index

- [User profile](user_profile.md) — Kien, solo AI builder, Python/TS
- [Project: magic-powers](project_magic_powers.md) — Claude Code plugin, v1.9.0
- [Preferences](feedback_coding.md) — uv not pip, direct comms, no over-engineering

Rule: Keep MEMORY.md under 200 lines or later lines get cut off at load time.

5. Project Continuity Pattern

At the end of significant sessions, update memory explicitly:

# project_context.md update (after a major session)

## Current State (updated 2026-04-08)
- Completed: MCP setup for postgres, github integrations
- In progress: Hook system for auto-testing on edit
- Blocked: Waiting for staging DB credentials from ops
- Next session: Wire up Stop hook to enforce pytest passing

This replaces the need to re-explain project state at the start of every session.

Key Concepts

• Auto-memory — Claude Code automatically saves important context to ~/.claude/projects/[hash]/memory/
• MEMORY.md — always-loaded index file; must stay under 200 lines
• Memory types — user (preferences), project (state), feedback (corrections), reference (links)
• Memory staleness — memories about code can go stale; always verify before acting on code-specific memories

Checklist

• MEMORY.md index exists and is under 200 lines?
• User profile memory set (role, communication style, key tools)?
• Project context memory updated when major decisions are made?
• Feedback memories added after correcting Claude's mistakes?
• Memory files verified as current (not stale from months ago)?

Key Outputs

• Memory directory structure with appropriate files per type
• MEMORY.md index with concise pointers to each memory file
• Project continuity entries capturing current state after major sessions

Output Format

• 🔴 Critical — no memory system (Claude restarts cold every session), MEMORY.md over 200 lines (truncated at load)
• 🟡 Warning — stale memories about code that no longer exists, no feedback memories (Claude repeats same mistakes)
• 🟢 Suggestion — add project memory after major architectural decisions, add user memory for communication preferences, update "Current State" after complex sessions

Anti-Patterns

• Storing code snippets in memory (code changes; reference files instead)
• MEMORY.md over 200 lines (truncated at load time — later entries are invisible)
• No feedback memories (Claude will repeat the same mistakes you've already corrected)
• Memory entries that say "as of [old date]" but were never updated — stale context misleads Claude

Integration

• Use with claude-md-authoring (CLAUDE.md = instructions; memory = accumulated context — they complement each other)
• Use with claude-project-settings (memory location is configured alongside other project settings)
• Agent: @technical-writer can help draft well-structured memory entries for complex projects