Memory

Standing instructions 9p reads at the start of every session.

Every session starts from zero unless you give it something to start from. 9P.md is that file.

Two files

File	Scope
~/.9p/9P.md	Global, applies in every project
./9P.md	Project, applies here, and wins where they disagree

Both are loaded automatically. Create them with:

9p init

What belongs in it

Standing instructions, the things you’d otherwise retype every session:

9P.md

# Project

Payments API. Node + Fastify + Postgres, deployed on Fly.

## Conventions
- Errors: return `Result<T, AppError>`, never throw across a module boundary
- Tests: Vitest, colocated as `*.test.ts`
- No default exports

## Workflow
- Run `npm run check` before saying a task is done
- Never edit `db/schema.sql` by hand. Write a migration

What works: conventions that aren’t obvious from the code, commands you always want run, and hard rules (“never do X”).

What doesn’t: restating what the code already says. That goes stale and 9p will trust it over the source.

Memory is capped at 8,000 characters. Past that it’s truncated, and a bloated 9P.md crowds out the actual conversation. Keep it tight.

Keeping it current

A useful line to include:

Update 9P.md before every commit if conventions changed.

Beyond a single file

For anything larger than standing instructions, decisions, per-feature notes, daily logs. Use a vault, which is a folder of linked markdown 9p can navigate rather than load wholesale.

Precedence

Global loads first, project second. Where they conflict, the project file wins by appearing later. Both are injected as standing instructions, not as conversation, so they survive /clear.