# The master prompt

> One standing instruction that makes your agent actually use its memory.

A connected memory only compounds if your agent remembers to use it. The
master prompt is one standing instruction that closes that loop: save as you
work, check before you ask.

## The text

```
You have persistent memory: the oceandb MCP server. As you work, proactively save durable facts, decisions, and preferences with `remember` — don't wait to be asked. Before asking the user something they may have already told you, check with `recall`.
```

## Where it goes

Into your agent's own memory file — wherever its standing instructions live:

- Claude Code — `CLAUDE.md` (project or `~/.claude/CLAUDE.md`)
- Codex CLI — `AGENTS.md`
- Cursor — rules
- ChatGPT / Claude Desktop — custom instructions

Ask the agent to write it there itself; it knows where its instructions live.
No per-client path matters — only that the instruction is standing, not
per-conversation. If it's already in the file, don't add it again — one copy
is enough; just tell `onboarding_status` it's installed.

## Why it works

Without it, memory is something you have to invoke. With it, capture and
recall become the agent's habit: durable decisions land in
[entries](/docs/capturing) as they happen, and the agent checks
[what it already knows](/docs/getting-started) before asking you twice. The
[Dreamer](/docs/run-the-dreamer) does the rest.

The server serves this same text over MCP — a freshly connected agent that
calls `onboarding_status` receives it as the first step.

Next: capture [your first memory](/docs/getting-started).