Auto Memory: Notes Claude Writes Itself

Auto Memory

Every time you start a new conversation with Claude, it begins with no memory of previous conversations. This is by design -- Claude does not automatically carry your context from session to session. Auto memory is the mechanism that changes this.

Auto memory is Claude's note-taking system. During a conversation, Claude can write observations about you -- your preferences, corrections you make, facts about your work -- to a persistent file on your machine. The next time you start Claude, those notes are loaded automatically. Claude already knows what it learned before.

Think of it like a colleague who keeps a notepad. Each time you work together, they check their notes from last time. When you correct something or state a preference, they write it down. Over time, they need less briefing and make fewer of the same mistakes.

How it works technically

Auto memory is stored in a file called MEMORY.md inside a folder that Claude manages on your machine. The location is:

~/.claude/projects/[your-workspace-path]/memory/MEMORY.md

The ~/.claude/ folder (that is, a folder called .claude in your home directory) is where Claude stores all of its local data. The projects/ subfolder contains one entry per workspace you have used. The MEMORY.md file inside each project folder is what Claude reads at the start of every session and writes to during sessions.

This file is stored entirely on your own computer. It is not sent to Anthropic's servers, it is not stored in the cloud, and it is not accessible to anyone else. You can open it, read it, edit it, or delete it using any text editor.

What gets remembered

Claude decides what to write to auto memory based on what seems worth preserving. In practice, common entries include:

• Explicit preferences -- things you state directly, such as "I prefer shorter paragraphs" or "always use UK English"
• Corrections -- when you tell Claude it got something wrong, or that its approach was not what you wanted
• Recurring context -- facts about your situation that come up repeatedly, such as "my business has four staff" or "I use Xero for accounting"
• Working conventions -- patterns Claude discovers during your sessions that are not yet written down anywhere, such as how you like documents formatted or what tone suits your audience

Claude does not write everything to memory -- only what it judges to be worth preserving. If you want something remembered, stating it clearly and explicitly helps. Vague feedback is less likely to be captured accurately than a direct statement of preference.

The 200-line limit

Auto memory files are capped at 200 lines. This limit exists to keep the file manageable -- a very long memory file would slow down session starts and consume context that could be used for actual work.

When the file approaches the limit, Claude begins condensing older entries to make room for new ones. Entries that are redundant, outdated, or less specific tend to be dropped or merged. This means the most recent and most frequently relevant information tends to survive.

If you notice that something is being dropped that you want kept, the right move is to put it in a CLAUDE.md file instead (see below). CLAUDE.md files have no size constraint and are under your direct control.

Auto memory vs CLAUDE.md: what is the difference

Feature	CLAUDE.md	Auto Memory (MEMORY.md)
Who writes it	You	Claude
When it changes	When you edit it manually	During sessions, automatically
Size limit	No limit	200 lines
Best for	Stable rules, tone, structure, non-negotiables	Evolving preferences, corrections, discovered patterns
Versioned in git	Usually yes	Usually no (gitignored)

Viewing your auto memory

Because auto memory is a plain text file, you can read it directly:

cat ~/.claude/projects/[workspace-path]/memory/MEMORY.md

You can also open it in any text editor -- it is a standard markdown file. In Claude Code, the /memory command opens an in-session browser that shows everything currently stored, organised by category.

Editing and resetting auto memory

You have full control over the file. To edit it, open it in a text editor and change whatever you want. To delete a single entry, remove that line. To start fresh, delete the file entirely -- Claude will create a new empty one at the start of the next session.

There is no hidden data. Everything Claude has recorded is visible in that file, in plain text.

What good auto memory looks like

Useful auto memory entries are specific and actionable. Compare these examples:

Weak entry	Strong entry
User likes clear writing	User prefers paragraphs of 2-3 sentences maximum. Corrected verbose responses twice.
Business context noted	User runs a 4-person HR consultancy serving SMEs in Ireland. Primary services: policy writing and training delivery.
Prefers UK English	Always use UK English spelling and grammar. User corrected "analyze" to "analyse" on 2026-03-12.

You can help Claude write better entries by being explicit when you give feedback. Instead of "that was a bit long", try "keep responses under 200 words for this type of task". The more specific the instruction, the more specific the memory entry.

When auto memory is not enough

Auto memory is good for preferences and context that develop organically. It is not the right tool for:

• Non-negotiable rules -- if there is something Claude must always do or never do, put it in CLAUDE.md. Auto memory can be overwritten; CLAUDE.md is under your direct control.
• Detailed instructions -- if a task requires a lengthy procedure (for example, a 12-step process for handling client enquiries), write it as a CLAUDE.md rule or a separate skill file. Auto memory entries are summaries, not procedures.
• Shared team setups -- auto memory is per-machine and per-user. If you need multiple people to share the same context, CLAUDE.md files committed to a shared repository are the right approach.
• Version-controlled preferences -- if you want a history of changes to your setup, CLAUDE.md in a git repository gives you that. Auto memory does not.