Append-only state as the unit of continuity

The problem

Work that spans more than one session needs a way to be picked up coherently by whatever runs next. The naive approach, a single state file the agent overwrites, loses history the instant a write goes wrong, and you can't reconstruct what the system believed before the bad write.

The shape

• Keep durable state as an append-only log under version control. New facts are added; old facts aren't silently rewritten.
• Commit every meaningful change with a message, so the history reads like a story you can follow.
• Keep one small, current summary on top for fast orientation, but treat the log beneath it as the source of truth.
• When something goes wrong, you roll back to a commit, not to a backup you hope still exists.

When to use it, and when not

Reach for this whenever continuity matters more than raw write speed: handoffs, long-running work, anything you'll need to audit later. Skip it for genuinely ephemeral scratch state, where the overhead of an audit trail buys you nothing. The test is simple: if future-you would ever ask "wait, when did it start believing that?", make it append-only.