lpm-memory

Project memory lives in one folder per project, shared by every agent CLI. Resolve that folder in this order:

1. $LPM_MEMORY_DIR, which lpm sets for every terminal it opens — always use it when set.
2. ~/.lpm/memory/<project> where <project> comes from lpm project --json: its parentName, or its name when parentName is empty.
3. Without the lpm CLI, ~/.lpm/memory/<project> where <project> is $LPM_PROJECT_NAME, else the main checkout's folder name (basename "$(dirname "$(cd "$(git rev-parse --git-common-dir)" && pwd)")", so a Git worktree resolves to the repository it belongs to), else the working directory's folder name.

A duplicate of a project shares the original's memory — same codebase, and the copy is disposable — so never derive a folder of its own from a copy's name or path. <session> is a kebab-case slug for one workstream (e.g. auth-refactor); each file is both the current handoff state and the work history.

Invocation:

• /lpm-memory <session-id> — Recall that session and continue it. Unknown id: offer close matches, or create it.
• No argument — Remember work already done in this conversation; at the very start of one, pick or create the session instead.
• Either way, keep the memory current from then on without being asked: append a timeline entry and refresh ## Current state after each milestone and when the user wraps up.

Remember (save / hand off)

1. Session slug: the user's name for it, else the one existing file that matches the work, else derive one and confirm it before writing.

2. Create the folder if missing. Seed a new file with:

   # <Work title>
   
   ## Goal
   <one or two lines>
   
   ## Current state
   <where things stand, next steps, blockers>
   
   ## Timeline

3. Re-read the file right before writing — another agent may have saved meanwhile. If it changed since you last read it, keep their changes: append your timeline entry after theirs and fold both realities into ## Current state.
4. Rewrite ## Current state to match reality now.

5. Append a new entry at the end of ## Timeline. The timeline is strictly append-only: never edit or delete an existing entry — not even your own from earlier in the same conversation. Each entry covers only what happened since the previous save:

   ### <YYYY-MM-DD HH:MM> — <agent>
   - Done: what shipped or changed, in outcome terms
   - Decided: choices made and why, including approaches tried and dropped
   - Learned: surprises and gotchas the next agent must know
   - Open: unresolved questions / blockers
   - Next: unfinished work / immediate next step

   <agent> = your CLI name (claude, codex, ...), local time. Drop empty lines; keep it brief.

6. Compaction — the one exception to append-only: when the timeline exceeds ten entries, condense the oldest into a single digest entry ### Archived through <YYYY-MM-DD> kept first in ## Timeline; keep the newest five entries verbatim. Condense only — preserve every decision and gotcha that still matters, never reinterpret, and fold into the existing digest on later compactions.

Recall (continue / join)

1. List the project memory folder's *.md; read the named session, or show the list (name, last modified, goal line) and ask which one.
2. ## Current state is the source of truth; the newest timeline entries carry the freshest detail. Read those first — reach for older entries and the archive digest only when the work needs that history.
3. State the next step you inferred, confirm direction, then continue — and Remember at the next stopping point.