Claude Memory & Claude Code Memory.

Does Claude have a memory?

Yes, on several levels in fact. Which one applies depends on whether you use Claude in the browser and the app, or as Claude Code in the terminal and your development environment.

Memory in the Claude app

The Claude app has a memory feature: Claude can search past chats and build a memory from them that flows into new conversations. You can turn it on, use it or turn it off in the settings, and saved memories can be deleted. Projects additionally bundle files and instructions for a topic area.

That's helpful for personal preferences and recurring topics. But for everyday work with code and agents, what matters most is how Claude Code remembers.

Claude Code memory: CLAUDE.md

The best-known form of Claude Code memory is the CLAUDE.md file. It lives in your project or your user folder and loads at the start of every session. Typical contents:

• commands for build, test and deployment,
• code conventions and architecture rules,
• reminders like "never push straight to main".

CLAUDE.md best practices call for short, clear files. Everything in it lands in the context of every session.

MEMORY.md and automatic memory files

Alongside CLAUDE.md, Claude Code now also writes its own notes to memory files like MEMORY.md. The idea: Claude remembers insights from the work without you writing them down. That's progress, but it has the same basic shape: a text file that gets loaded and grows over time.

Where memory files hit their limits

• All or nothing: The whole file is loaded, no matter what you're working on.
• Dilution: The longer it gets, the less weight each rule carries.
• No connections: A flat file doesn't know which rule belongs to which system.
• Outdated entries stay: Obsolete entries sit next to new ones, and contradictions pile up.
• Tied to one machine: What Claude Code learned on the laptop, the session on the server doesn't know.

Claude Code memory via MCP

With the Model Context Protocol (MCP), Claude Code can connect to external memory services. That gives Claude tools to store and retrieve memories. But many simple MCP memory servers wait for Claude to remember to query them. In the heat of work, that often doesn't happen.

Hebby: the long-term memory for Claude Code

Hebby goes further. It connects to Claude Code via hooks and MCP, runs automatically in every session and remembers on its own, without Claude having to ask. And it solves exactly the limits of memory files:

• Instead of all or nothing: only the memories that matter right now.
• Instead of a flat file: a graph memory that knows what belongs to which system.
• Instead of legacy clutter: what's outdated steps back on its own, and what proves itself gets stronger (Hebbian learning).
• Instead of one machine: one memory for all machines and agents, Codex included.

CLAUDE.md with Hebby: our recommendation

Long CLAUDE.md files cost tokens on every prompt and measurably make the agent worse (why). With Hebby, the rule is:

• Keep CLAUDE.md almost empty. Only what must hold without exception in every single session belongs in it. Usually that's a sentence or two.
• Turn off auto-memory. Built-in memory files can run in parallel, but they cost double. Hebby is built to take over that job.
• Hebby does the rest. Decisions, pitfalls and project knowledge arrive in the session on their own when they're needed.

Frequently asked questions