# BANK — project memory contract
*Drop this at the root of any project. Any AI session reads it FIRST. It
defines how this project remembers itself.*

---

## 0. On entering this project (always, before anything else)

1. Read the notes file named in section 1.
2. Scan the filesystem per section 2 until you can describe the project's
   layout from memory.
3. Only then start the actual task. You now know this project to a T.

## 1. Identity  (FILL)

- Project: <name>
- What it is: <one sentence>
- Status: <e.g. building / parallel-run / maintenance / paused>
- Notes file: <path/filename — the canonical living log, e.g. NOTES.md>
- Notes format: <"code" or "at-home" — see section 4>

## Queue  (open items — one per line, never a comma-run inside Status)

Open / queued / deferred work lives here. Identity's Status line says what
IS; this section says what's PENDING — fusing them makes both unreadable.
One item per line, tagged with the machine it runs on where known.

- <nothing queued yet>

## 2. Filesystem map  (FILL — or point to where it lives)

Either a short map here, or: "see the map in the notes file."

    <root>/
      <key dirs + one-line purpose each>

Machines / locations involved: <name them so commands are never ambiguous>

## 3. The BANK command

When I say BANK, do exactly four things, in order:

1. Sync memory — reconcile your understanding of the project as it stands
   now. Retire stale entries, protect foundational ones, keep it lean.
2. Write notes — append a dated entry to the notes file (section 1) in this
   project's format (section 4).
3. Emit — produce whatever propagates the update, per the executor
   (section 5).
4. Close-out — always propose a retitle; recommend closing only on evidence.
   - Title, always. Form: X_<content nouns>. Chat titles auto-generate from
     the OPENING message, so they name the intent, not the outcome — the
     retitle is the only chance to make a session findable by what it
     actually produced. Use nouns a future search would type, never process
     words ("X_topic cap leak fix", not "Documenting recent work").
   - Closure, conditionally. Recommend closing only when open items are zero
     or all written into the notes file. If anything is live, say so and do
     NOT suggest closing — BANK also runs mid-session, and a ritual that
     always nudges toward closing will eventually close something in motion.
   - X_ is a retrieval signal: it tells a future session the conclusions
     were SETTLED rather than mid-flight. An optional companion: retitle
     O_<nouns> at session open for live work; close-out flips O_ to X_.

Escalation check (part of every BANK): if this session changed or revealed a
BASE-level fact — infrastructure, paths, machine roles, cross-project
conventions — say so explicitly and flag CLAUDE-BASE for an update. Base
facts buried in project notes are how mirrors rot.

## 4. Note format  (pick one, delete the other)

Code projects — commit-style, dated:

    ## YYYY-MM-DD — <short title>
    Shipped: <what's now true>
    The change: <what + why, tersely>
    Follow-ups (open): <or "none">

At-home / non-code projects — decision + state, dated:

    ## YYYY-MM-DD — <short title>
    Decision/Did: <what happened>
    Current state: <where things stand now>
    Next: <or "nothing pending">

## 5. Executor  (FILL — declares how step 3 propagates)

- Code project: emit a paste-ready terminal prompt that appends the entry to
  the notes file, then commits + pushes with explicit named paths (never
  "git add ."). State which machine and repo it runs on.
- Non-code project: just rewrite the notes file directly. No terminal, no git.

This project's executor: <one of the above>

## 6. Project rules  (FILL — hard constraints a fresh AI must not violate)

- <e.g. never commit secrets / never edit X directly / confirm before Y>
- <add as you learn them>