diff --git a/src/simulator/crates/mc-ai/src/tactical/thresholds.rs b/src/simulator/crates/mc-ai/src/tactical/thresholds.rs
index 99ad03fd..5bb3d313 100644
--- a/src/simulator/crates/mc-ai/src/tactical/thresholds.rs
+++ b/src/simulator/crates/mc-ai/src/tactical/thresholds.rs
@@ -184,6 +184,30 @@ pub fn capital_walls_min_age_turns(axes: &BTreeMap<String, i32>) -> u32 {
     lerp_axis(defense, 30.0, 20.0, 10.0).round() as u32
 }
 
+// ─── Stateful-decisiveness thresholds (p1-29h) ─────────────────────────────
+
+/// Army-level commitment-hysteresis length, in turns, used by
+/// [`super::memory::TacticalMemory`]. Once the army commits to an objective it
+/// holds the lock for this many turns through tactical-score wobble (and the
+/// timer refreshes on each capture, so the army presses on rather than
+/// dispersing).
+///
+/// Mirrors the juiced harness `auto_play.gd::_attack_commitment_turns = 5`
+/// (`auto_play.gd:1146`) at the neutral axis, and lets personality lengthen or
+/// shorten it: aggressive + grudgeful clans stay committed longer (they finish
+/// what they start), cautious clans break off sooner.
+///
+/// Range: `composite=1` → 3 turns (cautious — re-evaluates often),
+/// `composite=5` → 5 turns (the `auto_play.gd` baseline),
+/// `composite=10` → 8 turns (relentless). Composite blends aggression 0.6 /
+/// grudge_persistence 0.4 — the same drivers as [`capital_approach_hex`].
+pub fn commitment_turns(axes: &BTreeMap<String, i32>) -> u32 {
+    let aggression = axis(axes, "aggression") as f32;
+    let grudge = axis(axes, "grudge_persistence") as f32;
+    let composite = ((aggression * 0.6 + grudge * 0.4).round() as i32).clamp(1, 10);
+    lerp_axis(composite, 3.0, 5.0, 8.0).round().max(1.0) as u32
+}
+
 #[cfg(test)]
 mod tests {
     use super::*;
@@ -224,6 +248,22 @@ mod tests {
         assert_eq!(final_push_enemy_city_count(&a), 1);
     }
 
+    #[test]
+    fn commitment_turns_baseline_matches_auto_play_constant() {
+        // axis=5 composite must reproduce auto_play.gd's _attack_commitment_turns = 5.
+        let a = axes(&[("aggression", 5), ("grudge_persistence", 5)]);
+        assert_eq!(commitment_turns(&a), 5);
+    }
+
+    #[test]
+    fn commitment_turns_scales_with_personality() {
+        let cautious = axes(&[("aggression", 1), ("grudge_persistence", 1)]);
+        let relentless = axes(&[("aggression", 10), ("grudge_persistence", 10)]);
+        assert!(commitment_turns(&cautious) < commitment_turns(&relentless));
+        assert!(commitment_turns(&cautious) >= 1, "never zero — always commit at least 1 turn");
+        assert_eq!(commitment_turns(&relentless), 8);
+    }
+
     #[test]
     fn capital_siege_no_retreat_hp_baseline_matches_historical() {
         let a = axes(&[("aggression", 5)]);
diff --git a/src/simulator/crates/mc-ai/src/tactical/tree_state.rs b/src/simulator/crates/mc-ai/src/tactical/tree_state.rs
index c30a02d3..969b200e 100644
--- a/src/simulator/crates/mc-ai/src/tactical/tree_state.rs
+++ b/src/simulator/crates/mc-ai/src/tactical/tree_state.rs
@@ -83,7 +83,11 @@ impl TreeState for TacticalTreeState {
         // Seed deterministically from depth so the candidate set is the
         // same on repeat traversals at the same depth.
         let mut rng = XorShift64::new(0x9E37_79B9_7F4A_7C15u64.wrapping_mul(self.depth as u64 + 1));
-        decide_tactical_actions(&self.inner, &self.scoring, &mut rng, None)
+        // p1-29h — MCTS rollouts are hypothetical lookahead with no cross-turn
+        // continuity, so each enumeration uses a fresh transient memory. The
+        // army-lock is a live-play decisiveness lever, not a search-tree one.
+        let mut memory = super::memory::TacticalMemory::default();
+        decide_tactical_actions(&self.inner, &self.scoring, &mut rng, None, &mut memory)
     }
 
     fn apply(&self, action: &Self::Action) -> Self {
@@ -110,8 +114,9 @@ impl TreeState for TacticalTreeState {
         // we don't depend on the tree's `root_player` config knob.
         let mut current = self.inner.clone();
         let _start = Instant::now();
+        let mut memory = super::memory::TacticalMemory::default();
         for _ in 0..horizon {
-            let actions = decide_tactical_actions(&current, &self.scoring, rng, None);
+            let actions = decide_tactical_actions(&current, &self.scoring, rng, None, &mut memory);
             if actions.is_empty() {
                 break;
             }