Auto-Memories: When and How

What is it?

Introduced in v2.1.59 (February 2026), auto-memories allow Claude Code to save useful context between sessions without manually editing CLAUDE.md. Claude identifies key decisions during conversations and stores them automatically for future sessions in the same project.

Do not confuse with Claude.ai memory (web interface). These are two distinct systems. Claude Code auto-memories are local, per-project, with no connection to your claude.ai account.

Where memories are stored

~/.claude/projects/<encoded-path>/memory/MEMORY.md

This is the native default. Configure autoMemoryDirectory only when you need a different location; a .claude/memory/ directory is not automatically the project memory store.

Since v2.1.63, auto-memories are shared across git worktrees of the same repo.

What gets memorized

Architectural decisions:
  "This project uses Prisma for database access"

Team preferences:
  "Functional components are preferred over classes"

Specific patterns:
  "API routes follow RESTful naming in /api/v1/"

Known issues:
  "Do not use package X, version conflict with Y"

CLAUDE.md vs auto-memories

Managing memories with /memory

# In Claude Code (interactive mode)
/memory          # View, edit, or delete entries

When to use which

Write in CLAUDE.md manually: team conventions, critical rules everyone must follow, official architectural decisions.

Let auto-memories work: personal workflow preferences, patterns discovered mid-session, temporary context tied to a feature in progress.

Review regularly: inspect memory with /memory and correct outdated notes. Auto memory is enabled by default and can be disabled with autoMemoryEnabled: false. The startup index is limited to 200 lines or 25KB; topic files load on demand. Default memory files are machine-local, not shared through the repository. Memory reference.