The Best Way to Give Claude Memory in 2026
There are three real ways to give Claude memory. Each one is right for a different situation. Here's an honest breakdown — no overselling, just the actual trade-offs.
The problem in one sentence
Claude is stateless by design. Every conversation starts from zero. Whatever it learned about you — your job, your projects, your preferences, your name — is gone the moment you close the tab.
This isn't a bug. It's a deliberate privacy choice. But it creates a real workflow problem for anyone using Claude regularly. You end up spending the first few minutes of every conversation re-briefing instead of working.
Here are your three real options.
Option 1: Custom instructions
The simplest fix. Claude has a "Custom instructions" field (Settings → Custom instructions) where you paste a paragraph or two about yourself. Claude reads it at the start of every conversation.
What it's good for
- Your identity and role ("I'm a software engineer at a B2B SaaS startup")
- Consistent tone preferences ("be direct, skip the preamble")
- A few standing facts ("I use Python, not JavaScript")
Where it breaks
Custom instructions have a hard character limit. Once you're trying to cram in role, projects, clients, preferences, and context, you're over the limit or you're writing something so compressed it loses specificity.
More importantly: it doesn't update. If you close a deal, finish a project, or change your priorities, you have to manually rewrite your instructions. Nothing saves automatically.
Verdict
Custom instructions are the right tool for stable, global facts about you. They break down for anything that changes or grows.
Option 2: Claude Projects
Claude's Projects feature (available on Pro) lets you create a persistent workspace with its own system prompt and file uploads. Everything in the Project carries across sessions within that Project.
What it's good for
- A focused, self-contained body of work (one report, one codebase)
- Reference documents you want Claude to have access to (a company brief, a style guide)
- Keeping multiple work streams separate
Where it breaks
Projects are siloed. Context from one Project doesn't flow into another. You can't search across Projects. And if you're working on something genuinely long-running — a job search, an ongoing client relationship, a research thread spanning months — Projects don't map cleanly.
The file upload approach also has a footprint problem: large documents bloat your context window on every message, which makes Claude slower and more expensive to run.
Verdict
Projects are the right tool for a bounded, document-heavy workstream. They're not a general memory layer.
Option 3: MCP (the external memory approach)
MCP — the Model Context Protocol — lets you connect Claude to external tools and databases. Instead of stuffing context into Claude's window, you give Claude the ability to look things up when it needs them.
This is architecturally different from the first two options. Custom instructions and Projects push data into the model. MCP gives the model a retrieval system — like the difference between memorising a book and having the book on your shelf.
What it's good for
- Large, growing bodies of information (hundreds or thousands of records)
- Data that changes regularly (client notes, project status, reading lists)
- Searching across everything you've ever saved
- Multiple Claude projects all sharing the same underlying data
Where it takes more setup
MCP connectors have to be added through Claude's settings. It's a URL paste, not a toggle — thirty seconds, but it exists. And you need something on the other end of the URL (a running MCP server). That's where something like Stash comes in: it's a hosted MCP server you don't have to build or run yourself.
Verdict
MCP is the right tool when you want Claude to have real, searchable, growing memory across all your work — not just within one Project or within a character limit.
The decision tree
| If you want... | Use |
|---|---|
| Claude to know your role and tone globally | Custom instructions |
| A bounded project with docs Claude should reference | Projects |
| Claude to remember everything, searchably, forever | MCP |
| All three at once | All three — they're not mutually exclusive |
The short answer: most people who use Claude seriously end up using all three. Custom instructions for the stable stuff. Projects for focused workstreams. MCP for the growing, searchable layer underneath everything.
Getting started with MCP memory
If you want to try the MCP approach, Stash is a hosted record store built specifically for Claude. You sign in with Google, get a connector URL, paste it into Claude's MCP settings, and you're done. No server to run, no API to configure.
Free tier covers 2,500 records and 50 searches a month — enough to get a sense of the pattern before committing. The connector URL is:
https://app.stashlite.com/mcpOnce it's connected, you tell Claude what to save. It searches when you need it. You never re-brief.