Turn Off Claude Code's Auto Memory and Steer It Yourself

Matt Pocock turns Claude Code's auto memory off completely because a stateless agent is more predictable than one carrying accumulated preferences into every session. In the video he opens the /memory command inside a WSL Claude Code session, points to the "Auto-memory: off" toggle, and argues that the feature saves too eagerly. His fix is to delete what the agent stores on its own and put any preference you actually want into a file you control, like CLAUDE.md or a doc in your repo.

What the video shows on screen

The screen recording is a VS Code window running Claude Code v2.1.205 on Opus 4.8 (1M context) in a repo called course-video-manager. Pocock runs the /memory slash command, described in the menu as "Open a memory file in your editor." The panel that opens is titled "Memory" and lists three things: an "Auto-memory: off" toggle, "Project memory" checked in at ./CLAUDE.md, and "User memory" saved in ~/.claude/CLAUDE.md. It also links the docs at https://code.claude.com/docs/en/memory. Pocock's point is anchored to that toggle. He wants it off.

Why he kills the memory

Pocock's objection is about behavior drift. Auto memory saves a file called MEMORY.md and uses it to remember your preferences across sessions. His complaint is that it is "way too freaking eager" to do so. A preference about how one specific feature should be built gets written to the memory bank, then leaks into every later conversation, whether or not it belongs there.

He contrasts that with a stateless agent. When the agent remembers nothing from previous sessions, he says, it is more predictable, because he works "from a consistent base every time." Memory buildup does the opposite: kludge accumulates over time, so every run is slightly different and slightly less predictable. His framing is about control. You should be steering, and if you have a preference you want expressed, you should express it deliberately rather than let it seep in from a saved note.

The alternative: files you write yourself

The recommendation is not to run without any persistent context. It is to keep the persistent context under your own hand. Pocock's line is to "edit your claw.md or make a little dock in your repo," meaning CLAUDE.md or a plain doc checked into the project. The distinction matters because those two mechanisms are different in kind. CLAUDE.md files are instructions you write. Auto memory is notes the agent writes itself from your corrections. Pocock trusts the first and switches off the second.

How this maps to the actual settings

Claude Code's documentation lines up with the panel in the video. Auto memory is on by default and can be toggled through the /memory command, which writes autoMemoryEnabled to your user settings. You can also set autoMemoryEnabled: false in a single project's settings, or disable it globally with the CLAUDE_CODE_DISABLE_AUTO_MEMORY=1 environment variable. Auto memory is stored per repository under ~/.claude/projects/<project>/memory/, with MEMORY.md as the index that loads into every session. That is the exact file Pocock is telling you to stop feeding.

Key Takeaways

Auto memory in Claude Code writes a MEMORY.md file and reuses it to recall preferences across sessions, and it is on by default.
Pocock turns it off entirely because it saves preferences too eagerly and lets narrow, one-off preferences influence unrelated conversations.
A stateless agent starts from a consistent base every session, which he finds more predictable than one carrying accumulated memory.
Toggle auto memory off with the /memory command, per-project via autoMemoryEnabled: false, or globally with CLAUDE_CODE_DISABLE_AUTO_MEMORY=1.
Keep persistent context in files you write yourself, CLAUDE.md or a repo doc, so you stay in control of what influences the agent.