Session Start

One job: load the right memory into context, cheaply. Nothing else.

A SessionStart hook (scripts/session-start-hook.sh) — or on pi, the Foundation extension running the same script — normally injects memory automatically — look for a === WORKING MEMORY === block in your context. If it's there, skip straight to step 6. Steps 1-5 are the manual fallback (hook not installed, or a harness with neither).

The hook injects MEMORY.md FIRST on purpose: it sits at the top of the prompt-cache prefix, so the harness caches it across the session. Changing MEMORY.md after the hook fires defeats this — it stays read-only until session end (see CLAUDE.md hard rules).

Steps

1. Read vault/MEMORY.md in full. This is the index — who you work for, active goals, standing lessons, and what Knowledge notes exist.
2. Read the most recent note in vault/Daily/ (today's if it exists, otherwise the latest). This tells you what happened last and what's open.
3. Read the most recent note in vault/Reflections/. Apply its lesson to how you work THIS session.
4. Do NOT bulk-read vault/Knowledge/. Open a Knowledge note only when the current goal needs that specific topic — the MEMORY.md index tells you which one.

5. If today's daily note doesn't exist, create vault/Daily/YYYY-MM-DD.md:

   # YYYY-MM-DD
   
   ## Sessions

6. Before starting work, state in one short block:
   • The goal as you understand it
   • Which open items from the daily note are relevant
   • Any assumptions you're making — if the goal is ambiguous, ask instead of guessing

Rules

• This skill loads memory. It never writes Knowledge or Reflections.
• Never re-read files the hook already injected — that's paying twice.
• Total reading: MEMORY.md + 1 daily note + 1 reflection + at most 2-3 Knowledge notes. If you're reading more than that, stop — you're burning context the session needs for actual work.