How to Save Context in Claude
Projects, memory, and MCP - three ways to stop re-explaining yourself
Every new Claude conversation starts fresh. It doesn't know your role, your projects, or your preferences. Re-explain everything, or get generic answers.
Claude has three ways to fix this. Here's what each does, how to set it up, and where it falls short.
Option 1: Profile Preferences
Whatever you write here loads into every conversation, every project, everywhere. The baseline layer.
How to set it up:
- Click your profile icon (bottom left on claude.ai)
- Settings
- Find "What preferences should Claude consider in responses?"
- Write your instructions and save
What to put there:
- Your role and professional context
- Communication style ("direct, no preamble," "skip caveats unless they matter")
- Standing constraints ("I work in B2B SaaS," "my audience is non-technical")
- Anything that applies across everything you use Claude for
The limit: 1,500 characters - roughly 250 words. Enough for the essentials, not deep context. Everything here loads into every conversation, so task-specific instructions belong in Projects, not here.
Option 2: Project Instructions
Projects are persistent workspaces. Each has its own instructions, document library, and conversation history. Instructions set at the project level apply to every conversation inside it.
How to set it up:
- Left sidebar → Projects → New Project
- Name it
- Click "Set project instructions"
- Add what's specific to this project: the tech stack, the client, the goal, the constraints
- Upload relevant documents to the knowledge base
How it stacks: both layers load. Profile Preferences set the baseline. Project instructions add specifics on top. No need to repeat global preferences in every project.
What to put there:
- Project-specific background ("Next.js app using Supabase. Follow existing patterns.")
- Role-specific context for this work ("writing for a technical audience, not executives")
- Documents: style guides, specs, briefs, reference material - up to 200K tokens per project
The limit: ~8,000 characters for instructions, plus the knowledge base. Available on all plans; free users get up to 5 projects.
Option 3: Styles
Separate from instructions. Styles control how Claude formats and delivers responses - switch per conversation without touching preferences.
Built-in options: Concise, Explanatory, Formal. You can create custom styles from your own writing samples.
Not where personal context lives. This is about response format, not who you are or what you're working on.
How the layers fit together
| Layer | Scope | Limit | Best for |
|---|---|---|---|
| Profile Preferences | Every conversation | 1,500 chars | Role, style, standing constraints |
| Project Instructions | One project | ~8,000 chars + docs | Project-specific context and documents |
| Styles | Per conversation | - | Response format and voice |
They load in order: Profile Preferences → Project Instructions → Styles. Later layers add specificity without overriding what came before.
What doesn't work well
Relying on conversation history: Claude doesn't carry memory between separate conversations, even inside a project. Within a conversation it remembers everything. Start a new one: blank slate, except for your instructions and documents.
Overloading Profile Preferences: 1,500 characters fills up fast. Too much dilutes the most important instructions. Keep it to what's genuinely universal.
Outdated instructions: stale context is often worse than no context. Profile Preferences that describe a project you finished six months ago get applied confidently to everything you do now. Review regularly.
The limitation instructions can't fix
Profile Preferences and Project Instructions store your context inside Claude. Fine when Claude is your primary tool.
Most people use more than one AI tool. The preferences you've set in Claude don't exist in Cursor. The project instructions you've built don't carry to ChatGPT. Every new tool starts from zero.
Structural limit of platform-specific context: each platform owns it.
The alternative is context that lives outside any platform - a structured profile you own, served to whichever tools you authorize. Connect a new tool and it already knows who you are.
→ How that works: What Is Personal Context for AI?
→ Set it up in Claude via MCP: How to Load Your Personal Context into Claude and Cursor via MCP
