docs: mark off-screen AW memory as shipped in architecture §5

docs/keel-architecture.md

@@ -185,8 +185,9 @@ mode would just route a thought. The runtime state machine becomes the loop in
 
 Off-screen mode, from the phone, end to end — **this is what ships today**:
 1. **Collect** — today's Marvin tasks (`am --json`) + the `~/owc/goals-2026.md`
-   goals + the life-domain `~/owc/resources/bug-*.md` notes, assembled into a
+   goals + the life-domain `~/owc/resources/bug-*.md` notes — plus recent
-   single budgeted brief.
+   proposals and their outcomes recalled from the `keel.events` AW bucket —
+   assembled into a single budgeted brief.
 2. **Brain** — the chosen brain (`ai.Proposer`) returns one worthwhile off-screen
    action plus a short rationale.
 3. **Surface** — the phone-first web UI renders a proposal card and asks Felix to
@@ -196,10 +197,13 @@ Off-screen mode, from the phone, end to end — **this is what ships today**:
 
 No new store touched. The brain was rented for one call. Everything else was Keel.
 
-**Next increment:** memory. Off-screen does not yet read or write Keel's AW
+**Shipped:** memory. Off-screen now writes its decisions to the `keel.events` AW
-buckets, so it cannot remember what it proposed last time. Adding a *remember*
+bucket (`proposal_made` / `action_taken` / `proposal_dismissed`, correlated by a
-step (read `keel.state` for prior proposals) and an *act* event
+random `proposal_id`) and reads its recent history back into the brief — so a
-(`proposal_made`/`action_taken` to `keel.events`) closes the loop per §7.
+proposal can avoid repeating itself and follow up on what was dismissed or left
+undone. Storage sits behind a thin `memory.Store` port (AW-backed, or a nop when
+AW is down), the seam every later mode reuses; see §7. The `keel.state` bucket and
+the trend/spiral detection that reads this data remain future increments.
 
 ---