How to Make Claude Remember Your Preferences Across Conversations

"Use uv, not pip." "Bullet points, not essays." "Always run tests before committing." How many times have you typed the same preferences to Claude before it finally sticks? Claude is excellent within a conversation and frustratingly forgetful between them.

Here is every practical way to make your preferences persist, from built-in features to skills to a shared memory layer.

What Claude remembers on its own

Claude has built-in memory features that store preferences and context across conversations, plus a memory import tool for bringing context over from other AI platforms. Check what it has stored and curate it occasionally; stale preferences ("we use library X" after you migrated to Y) are worse than none.

Claude also has Projects: dedicated workspaces where related chats and files stay grouped, and context carries across sessions within the project. Projects solve memory within a body of work, not across your life.

Method 1: CLAUDE.md (and the global one)

For Claude Code users, CLAUDE.md is the highest-leverage memory file you own:

• Project CLAUDE.md: lives in the repo, holds coding standards, architecture decisions, conventions. The team shares it.
• Global ~/.claude/CLAUDE.md: applies to every project. This is where personal preferences belong: communication style, tool choices ("use uv not pip"), workflow rules ("run tests before commit").

Keep it specific ("Use Inter, 4px radius, zinc-800 backgrounds" beats "make it look good") and organized with headers. Review it monthly; it rots.

Method 2: Memory skills

The Claude ecosystem has skills that teach Claude to detect and store preferences automatically. The pattern: when you correct Claude ("actually, use X not Y"), the skill saves it to local vector memory, and before responding in future sessions, Claude semantically checks that memory for relevant preferences. Some implementations are fully local with no cloud involved. This is the closest thing to "it just learns me" without a service.

Method 3: Memory protocol files

A related DIY approach: maintain a ~/.claude/memory.md (or similar) with sections for identity, preferences, projects, past decisions, and people, and instruct Claude to read and update it. Structured, transparent, fully yours. The cost is maintenance: you are the database administrator of your own preferences.

The maintenance problem

Every file-based method converges on the same chore: you keep the files accurate. Preferences change, projects end, decisions get reversed. A 400-line memory file full of stale entries is a liability, and curating it feels like a second job. That administrative burden is what pushes people toward managed options.

What I actually use: preferences that maintain themselves

I wanted the skill-style behavior (automatic detection, semantic recall) without maintaining files, and I wanted it everywhere, not just in Claude.

So I connected Vilix AI over MCP. It is a memory layer outside any single app:

• Automatic capture: corrections and preferences are saved from the conversation itself. No "remember this" commands, no file editing.
• Retrieval before reply: when a new conversation starts, relevant preferences surface on their own.
• Cross-tool: the same preferences apply in Claude Code, Cursor, and my other connected tools, on my laptop and phone, because the memory is stored server-side in my account.
• Setup is ~10 minutes per tool, and there is a free tier.

The difference in daily life is small but constant: I stopped prefacing requests with preferences I have stated a dozen times before.

Honest limits

Automatic memory is probabilistic. For hard rules that must always apply ("never deploy on Fridays"), keep them in CLAUDE.md or project instructions where they are deterministic. Use the memory layer for the long tail of preferences, context, and history that you should never have to repeat. And glance at what is stored now and then; every memory system benefits from a curator, even a lazy one.

The short version

Method	Automatic	Cross-tool	Maintenance
Built-in memory + Projects	Partly	No	Low
CLAUDE.md (global)	No	No	You
Memory skills	Yes	No	Low-Medium
Memory protocol files	No	No	You
Vilix AI	Yes	Yes	Minimal

If Claude is your only tool, global CLAUDE.md plus a memory skill gets you far. If your preferences should follow you everywhere, use one memory for all of it.