Every Claude conversation starts from zero. Claude doesn't know your name, your job, your ongoing projects, or what you talked about last week. For a one-off task that's fine. As a personal assistant you use every day, it's exhausting.
The good news: there are three layers you can configure that change this — and most Claude Pro users have only set up one of them.
Claude's custom instructions let you describe who you are so Claude doesn't need to ask. Every conversation gets this context automatically.
What to put there:
Custom instructions are good for identity and style. They are a poor fit for anything that changes — projects, tasks, contacts, notes. The limit is 1,500 characters, and trying to cram a project brief or contact list in there is a losing game.
Claude's Projects feature lets you attach files and instructions to a specific project — a long document, a style guide, a dataset. When you open a chat inside that project, the attached context is there.
Projects are the right tool when:
They're not the right tool when you have information that accumulates over time — notes from fifty conversations, a list of 300 contacts, a year's worth of decisions. That stuff doesn't fit in a Project attachment. You end up either truncating it or loading more tokens than the task needs.
This is where most people's setups stop. But there's a third layer that unlocks something neither custom instructions nor Projects can do: a searchable persistent store that grows with you and returns only what you need, when you need it.
That's what an MCP (Model Context Protocol) connector provides. Claude connects to an external record store. Instead of front-loading every bit of context at conversation start, you ask Claude to retrieve what's relevant. The rest stays on the server — zero token cost until it's needed.
The practical difference looks like this:
Claude retrieved one record. It didn't load your entire project history. It didn't need you to paste anything in.
If you're setting this up fresh, here are the five collections worth creating immediately. Each one compounds over time.
A short brief on who you are, what you're working on, and what matters right now. Updated monthly. Called with context() to prime any new session without repeating yourself. This is the most used collection in most setups.
Anything worth keeping from a conversation: a phrase that clicked, a decision rationale, a half-formed idea. Stored mid-conversation, searchable forever. You stop worrying about losing good thinking because saving takes one sentence.
Articles and links you intend to read. Instead of fifty open tabs or a bookmark graveyard, you ask Claude to save it. Retrieve by topic, tag, or keyword in any future session.
A lightweight CRM. "Save: Sarah Chen, head of design at Meridian, met at UX London, intro via Marcus. Follow up about the research internship." Retrieve with "who do I know at Meridian?" or "who am I supposed to follow up with this week?"
A record of choices and the reasoning behind them. When you're six months into a project and can't remember why you chose one database over another, you ask Claude. It finds the decision note with the date and your reasoning. This pays back faster than people expect.
Once you have standing context stored, the morning routine becomes a single line:
That's a cold start into a warm conversation. No preamble, no re-explaining. The context hit costs a fraction of a cent and Claude is immediately useful.
Honest constraints — you should know these before you commit to any tool:
Stash is a hosted MCP server. There's nothing to install or run. The setup is:
Your account is provisioned automatically. A seeded standing-context collection is ready to fill in.
| Layer | What it stores | How it's loaded | Scales? |
|---|---|---|---|
| Custom instructions | Identity, style, preferences | Auto, every conversation | No — 1,500 char cap |
| Projects | Reference docs, stable context | Auto, within the project | Partially — token cost grows with attachment size |
| MCP connector (Stash) | Notes, contacts, decisions, anything that grows | On demand via tool call | Yes — only retrieves what's asked for |
None of these replaces the others. Custom instructions set the baseline tone. Projects handle specific reference material. The MCP layer handles everything that accumulates over time.
Most Claude Pro users have only done Layer 1. Adding Layer 3 is where the real leverage is.
Add the persistent memory layer
Free tier. Sign in with Google. 30 seconds to set up.
Get your connector URL →Stash is a new service. Free tier limits and paid pricing (currently £8/month for Pro) may change. Cancel anytime.