Pick Up a Multi-Week Project Exactly Where You Left Off

Dozens of sessions, one project, zero re-explaining. How goals, locked decisions, and next steps carried a multi-week build across every session that touched it.

The problem: long projects, short-lived context

Session continuity is the thing AI coding tools are worst at. Claude Code, Cursor, whatever you use -- a session's context dies with the session. On a one-afternoon task, that's fine. On a multi-week project, it's brutal: every new session opens on a project it has never seen, and you burn the first stretch of it rebuilding context by hand.

Worse than the lost minutes is the lost direction. A fresh session doesn't just forget what's next -- it forgets what was already settled. It will happily re-open questions you closed a week ago, propose the approach you rejected, and "improve" things a past session deliberately chose not to do.

The real story: a plan made Tuesday, executed days later by a stranger

This is our own workflow. One of our projects was a growth-engine build that ran for weeks and spanned dozens of sessions -- different days, different chats, sometimes different machines.

Three kinds of memory carried it:

• Goals -- what the project is trying to achieve, set once and updated rarely
• Locked decisions -- calls that were made deliberately, marked "don't re-litigate," each with its reasoning attached
• Next steps -- the concrete queue of what to do, written down the moment it's known

All three auto-load at session start. So when we finalized a plan on a Tuesday and didn't get back to it for days, the session that eventually executed it was a completely fresh session -- a stranger to the whole project -- and it executed the plan correctly, with zero re-explaining. It opened, read the goals and the queue, and went to work.

And because decisions carry their why, later sessions didn't undo them. A session that can see "we chose X over Y, here's the reasoning" doesn't wander back to Y. A session that can only see the code does, constantly.

How it works

memoir gives your AI tool memory over MCP. During a session, the assistant records the state of the project as it changes:

# Set the destination once
memoir_set_goal("Ship the growth engine: automated
  posting + a feedback loop from engagement data")

# Queue work the moment it's decided
memoir_add_next("Wire the scoring report into the
  daily run -- plan finalized, see notes")

# Lock a decision, with the why
memoir_note("decisions", "Deploy from main, not from
  feature branches. LOCKED -- don't re-litigate.
  Why: a worktree deploy once shipped stale code.")

Then the start of every future session looks like this, with no one asking for it:

# A fresh session, days later

Goals:  Ship the growth engine: automated posting +
        a feedback loop from engagement data
Next:   - [ ] Wire the scoring report into the daily
              run -- plan finalized, see notes

The new session starts where the last one stopped -- same goals, same queue, same settled decisions. When it needs the deeper reasoning, it calls memoir_recall and reads the full note.

More use cases

• Make your AI remember between sessions -- a reminder that surfaced on the right day, sessions later
• An AI assistant that learns from its mistakes -- one production incident, recorded once, never repeated
• Every decision, with the why attached -- a decision log your AI writes as decisions happen