diff --git a/src/openhuman/memory_sync/composio/periodic.rs b/src/openhuman/memory_sync/composio/periodic.rs
index b7da0f811..34f66465b 100644
--- a/src/openhuman/memory_sync/composio/periodic.rs
+++ b/src/openhuman/memory_sync/composio/periodic.rs
@@ -52,7 +52,10 @@ use tokio::time::interval;
 
 use crate::openhuman::config::rpc as config_rpc;
 use crate::openhuman::config::DEFAULT_MEMORY_SYNC_INTERVAL_SECS;
-use crate::openhuman::scheduler_gate::gate::current_policy;
+use crate::openhuman::memory_sources::{
+    memory_sync_defaults_for_toolkit, MemorySourceEntry, SourceKind,
+};
+use crate::openhuman::scheduler_gate::gate::{current_policy, resume_notify};
 use crate::openhuman::scheduler_gate::policy::PauseReason;
 
 use super::providers::{get_provider, ComposioUsage, ProviderContext, SyncReason};
@@ -192,6 +195,57 @@ fn persisted_since_last_sync(
     })
 }
 
+/// Outcome of consulting the per-source registry for one Composio connection
+/// during a periodic tick (#2831).
+#[derive(Debug, PartialEq, Eq)]
+enum PeriodicSourceDecision {
+    /// The user toggled this source **off** — skip background sync entirely.
+    /// Manual `memory_sources_sync` still works (it has its own `enabled`
+    /// guard); only the automatic loop honours this here.
+    Skip,
+    /// Sync this connection with the given caps (`None` = uncapped for that
+    /// dimension).
+    Sync {
+        max_items: Option<u32>,
+        sync_depth_days: Option<u32>,
+    },
+}
+
+/// Decide whether — and with what caps — to periodically sync one connection,
+/// honouring the per-source `enabled` toggle (#2831). Pure so the three
+/// branches can be unit-tested without async/registry I/O.
+///
+/// - **disabled row** → [`PeriodicSourceDecision::Skip`]: the background loop
+///   must not sync a source the user switched off (this was the row-2 leak —
+///   previously the loop only read the registry for caps and synced disabled
+///   sources anyway, uncapped).
+/// - **enabled row** → `Sync` with the row's caps.
+/// - **no row yet** → `Sync` with conservative per-toolkit defaults
+///   ([`memory_sync_defaults_for_toolkit`]). `reconcile` normally backfills an
+///   enabled, capped row for every connection; this covers the brief
+///   pre-reconcile window. The no-match path defaults to *sync-bounded*, never
+///   *skip* and never *uncapped*, so a missing/mismatched row degrades safely
+///   (data keeps flowing, just capped) instead of silently going dark.
+fn decide_periodic_source(
+    source: Option<&MemorySourceEntry>,
+    toolkit: &str,
+) -> PeriodicSourceDecision {
+    match source {
+        Some(s) if !s.enabled => PeriodicSourceDecision::Skip,
+        Some(s) => PeriodicSourceDecision::Sync {
+            max_items: s.max_items,
+            sync_depth_days: s.sync_depth_days,
+        },
+        None => {
+            let (max_items, sync_depth_days) = memory_sync_defaults_for_toolkit(toolkit);
+            PeriodicSourceDecision::Sync {
+                max_items,
+                sync_depth_days,
+            }
+        }
+    }
+}
+
 /// Spawn the periodic sync background task. Idempotent: only the
 /// first call actually spawns the loop, every subsequent call is a
 /// cheap no-op (logged at `debug` so it's visible during startup
@@ -221,14 +275,35 @@ pub fn start_periodic_sync() {
 
 /// Inner loop, broken out so it's easy to mock-replace in tests if we
 /// ever want to drive ticks deterministically.
+///
+/// Each iteration waits on whichever comes first (#2831):
+///   * the 20-min `ticker` — the steady-state cadence, or
+///   * the scheduler-gate **resume** notify — fired when the user toggles
+///     Memory Tree back on or signs back in.
+///
+/// On a resume wake we run a tick **immediately** (so sync restarts within
+/// seconds, not at the next ≤20-min boundary) and `reset()` the ticker so the
+/// *next* scheduled tick is a full `TICK_SECONDS` out. The reset is what stops
+/// rapid off-on toggling from bunch-firing: many wakes collapse into at most
+/// one extra tick (the `Notify` stores a single permit), and the cadence
+/// re-bases from the last actual tick.
 async fn run_loop() {
     let mut ticker = interval(Duration::from_secs(TICK_SECONDS));
+    let resume = resume_notify();
     // Skip the immediate-fire tick so startup isn't slammed before the
     // user even has time to sign in.
     ticker.tick().await;
 
     loop {
-        ticker.tick().await;
+        tokio::select! {
+            _ = ticker.tick() => {}
+            _ = resume.notified() => {
+                // Woke early on a resume transition. Re-base the cadence so the
+                // next scheduled tick is TICK_SECONDS from now, then fall
+                // through and run the tick immediately.
+                ticker.reset();
+            }
+        }
         if let Err(e) = run_one_tick().await {
             tracing::warn!(
                 error = %e,
@@ -378,6 +453,28 @@ pub(crate) async fn run_one_tick() -> Result<(), String> {
     let audit_index = index_last_success_by_connection(&read_audit_log(&config));
     let now = Utc::now();
 
+    // Per-source registry snapshot (#2831). The periodic loop gates on the
+    // per-source `enabled` toggle so a source the user switched off stops
+    // syncing in the background — matching the manual paths
+    // (`memory_sources::sync_source`, `memory_sources_sync_all`), which already
+    // early-return on `!enabled`. Index every Composio source (enabled and
+    // disabled) by connection id; the per-connection branch below resolves
+    // skip/caps via `decide_periodic_source`.
+    //
+    // Built from the **already-loaded** `config` snapshot (Step 1), not a second
+    // `list_sources()` read. A separate read whose error we swallowed to an
+    // empty map would make every disabled source fall through to the
+    // `decide_periodic_source(None, ..)` default-caps path — silently
+    // re-enabling background sync for sources the user switched off on a
+    // transient config-read failure. Reusing the tick's snapshot is fail-closed
+    // (a disabled row stays disabled) and avoids the extra read entirely.
+    let composio_sources: HashMap<String, MemorySourceEntry> = config
+        .memory_sources
+        .iter()
+        .filter(|s| s.kind == SourceKind::Composio)
+        .filter_map(|s| s.connection_id.clone().map(|id| (id, s.clone())))
+        .collect();
+
     let mut considered = 0usize;
     let mut fired = 0usize;
     for conn in resp.connections {
@@ -425,20 +522,25 @@ pub(crate) async fn run_one_tick() -> Result<(), String> {
             continue;
         }
 
-        // Look up per-source caps from the memory_sources registry.
-        // Non-fatal: if the lookup fails we proceed without caps.
-        let (src_max_items, src_sync_depth_days) = {
-            let registry_sources = crate::openhuman::memory_sources::list_enabled_by_kind(
-                crate::openhuman::memory_sources::SourceKind::Composio,
-            )
-            .await
-            .unwrap_or_default();
-            registry_sources
-                .iter()
-                .find(|s| s.connection_id.as_deref() == Some(conn.id.as_str()))
-                .map(|s| (s.max_items, s.sync_depth_days))
-                .unwrap_or((None, None))
-        };
+        // Per-source gate + caps from the memory_sources registry (#2831).
+        // A disabled source is skipped here (the background-sync half of the
+        // toggle); enabled sources sync with their caps; a connection with no
+        // registry row yet syncs with conservative per-toolkit defaults.
+        let (src_max_items, src_sync_depth_days) =
+            match decide_periodic_source(composio_sources.get(&conn.id), &toolkit) {
+                PeriodicSourceDecision::Skip => {
+                    tracing::debug!(
+                        toolkit = %toolkit,
+                        connection_id = %conn.id,
+                        "[composio:periodic] source disabled — skipping periodic sync"
+                    );
+                    continue;
+                }
+                PeriodicSourceDecision::Sync {
+                    max_items,
+                    sync_depth_days,
+                } => (max_items, sync_depth_days),
+            };
 
         tracing::debug!(
             toolkit = %toolkit,
@@ -633,6 +735,126 @@ mod tests {
         ));
     }
 
+    /// Build a minimal Composio `MemorySourceEntry` for the per-source gate
+    /// tests — only the fields `decide_periodic_source` reads are meaningful.
+    fn composio_source(
+        enabled: bool,
+        max_items: Option<u32>,
+        sync_depth_days: Option<u32>,
+    ) -> MemorySourceEntry {
+        MemorySourceEntry {
+            id: "src_test".to_string(),
+            kind: SourceKind::Composio,
+            label: "test".to_string(),
+            enabled,
+            toolkit: Some("gmail".to_string()),
+            connection_id: Some("cmp-1".to_string()),
+            path: None,
+            glob: None,
+            url: None,
+            branch: None,
+            paths: Vec::new(),
+            max_commits: None,
+            max_issues: None,
+            max_prs: None,
+            query: None,
+            since_days: None,
+            max_items,
+            selector: None,
+            max_tokens_per_sync: None,
+            max_cost_per_sync_usd: None,
+            sync_depth_days,
+        }
+    }
+
+    /// #2831 row 2: a source explicitly toggled **off** must be skipped by the
+    /// background loop — this is the leak the gate closes.
+    #[test]
+    fn decide_periodic_source_skips_disabled_source() {
+        let src = composio_source(false, Some(100), Some(30));
+        assert_eq!(
+            decide_periodic_source(Some(&src), "gmail"),
+            PeriodicSourceDecision::Skip
+        );
+    }
+
+    /// An enabled source syncs with exactly its configured caps (no defaulting).
+    #[test]
+    fn decide_periodic_source_uses_enabled_source_caps() {
+        let src = composio_source(true, Some(42), Some(7));
+        assert_eq!(
+            decide_periodic_source(Some(&src), "gmail"),
+            PeriodicSourceDecision::Sync {
+                max_items: Some(42),
+                sync_depth_days: Some(7),
+            }
+        );
+    }
+
+    /// A connection with no registry row yet (pre-reconcile window) syncs with
+    /// the conservative per-toolkit defaults — **bounded**, never uncapped, and
+    /// never skipped. This is the safe-direction fallback for a missing match.
+    #[test]
+    fn decide_periodic_source_defaults_caps_when_no_row() {
+        let (want_items, want_depth) = memory_sync_defaults_for_toolkit("gmail");
+        assert_eq!(
+            decide_periodic_source(None, "gmail"),
+            PeriodicSourceDecision::Sync {
+                max_items: want_items,
+                sync_depth_days: want_depth,
+            }
+        );
+        // The defaults are bounded for a known toolkit (regression guard against
+        // an accidental return to uncapped background fetches).
+        assert!(want_items.is_some());
+    }
+
+    /// Multi-account regression (#3443 added multiple account connections per
+    /// toolkit): two live connections of the *same* toolkit must be gated
+    /// **independently** by their own per-`connection_id` source rows. This
+    /// pins the loop's connection-id keying — re-keying the lookup by toolkit
+    /// would collapse the two accounts and is the regression this guards.
+    #[test]
+    fn per_connection_gate_is_independent_across_accounts_of_same_toolkit() {
+        // gmail account A: enabled with caps; gmail account B: disabled.
+        let mut a = composio_source(true, Some(10), Some(5));
+        a.connection_id = Some("conn-A".to_string());
+        a.toolkit = Some("gmail".to_string());
+        let mut b = composio_source(false, Some(99), Some(99));
+        b.connection_id = Some("conn-B".to_string());
+        b.toolkit = Some("gmail".to_string());
+
+        // Build the same connection_id → entry index the live tick builds.
+        let index: HashMap<String, MemorySourceEntry> = [a, b]
+            .into_iter()
+            .filter_map(|s| s.connection_id.clone().map(|id| (id, s)))
+            .collect();
+
+        // Account A (enabled) syncs with its own caps...
+        assert_eq!(
+            decide_periodic_source(index.get("conn-A"), "gmail"),
+            PeriodicSourceDecision::Sync {
+                max_items: Some(10),
+                sync_depth_days: Some(5),
+            }
+        );
+        // ...account B (disabled) is skipped, even though it shares the toolkit.
+        assert_eq!(
+            decide_periodic_source(index.get("conn-B"), "gmail"),
+            PeriodicSourceDecision::Skip
+        );
+        // A third, not-yet-registered account of the same toolkit falls back to
+        // bounded defaults (never skipped, never uncapped).
+        let (def_items, def_depth) = memory_sync_defaults_for_toolkit("gmail");
+        assert_eq!(
+            decide_periodic_source(index.get("conn-C"), "gmail"),
+            PeriodicSourceDecision::Sync {
+                max_items: def_items,
+                sync_depth_days: def_depth,
+            }
+        );
+    }
+
     /// End-to-end simulation of the scheduler's per-connection decision: prove
     /// that **changing the global setting changes when the next sync fires**
     /// (issue #3302 acceptance criterion). We drive the same two pure helpers
diff --git a/src/openhuman/scheduler_gate/gate.rs b/src/openhuman/scheduler_gate/gate.rs
index 96fd7a846..728893d1e 100644
--- a/src/openhuman/scheduler_gate/gate.rs
+++ b/src/openhuman/scheduler_gate/gate.rs
@@ -10,7 +10,7 @@ use std::sync::{Arc, OnceLock};
 use std::time::Duration;
 
 use parking_lot::RwLock;
-use tokio::sync::{OwnedSemaphorePermit, Semaphore};
+use tokio::sync::{Notify, OwnedSemaphorePermit, Semaphore};
 
 use crate::openhuman::config::{Config, SchedulerGateConfig};
 use crate::openhuman::scheduler_gate::policy::{decide, PauseReason, Policy};
@@ -240,12 +240,50 @@ pub fn init_global(config: &Config) {
     });
 }
 
+/// Process-wide resume signal (#2831). Fired whenever the gate transitions
+/// **out of** a paused state — the user toggles Memory Tree back on
+/// ([`update_config`]) or signs back in ([`set_signed_out`]). Background loops
+/// (e.g. the Composio periodic scheduler) park on [`resume_notify`] so they can
+/// resume work within seconds instead of waiting out their next tick boundary.
+static RESUME_NOTIFY: OnceLock<Arc<Notify>> = OnceLock::new();
+
+/// Handle to the process-wide resume [`Notify`] (#2831).
+///
+/// Both the firing side (`update_config` / `set_signed_out`) and the waiting
+/// side (background loops) call this, so they share one instance. Use
+/// `notify_one()` to fire: if a loop is parked it wakes immediately; if it's
+/// mid-tick, a single permit is stored so the *next* `notified()` returns at
+/// once — a resume that arrives during a tick is never lost.
+///
+/// **Over-notifying is safe by design.** A spurious wake (e.g. Memory Tree
+/// toggled on while still signed out, so the effective policy is still paused)
+/// just causes one cheap gate-checked tick that re-reads [`current_policy`] and
+/// no-ops. We therefore fire on each individual un-pause transition rather than
+/// computing the precise combined (config × signed-out) edge.
+pub fn resume_notify() -> Arc<Notify> {
+    RESUME_NOTIFY
+        .get_or_init(|| Arc::new(Notify::new()))
+        .clone()
+}
+
 /// Update the gate's view of user config (e.g. after a settings change).
+///
+/// Fires [`resume_notify`] when this update moves the policy out of a paused
+/// state (e.g. Memory Tree toggled back on), so parked background loops resume
+/// promptly (#2831).
 pub fn update_config(cfg: SchedulerGateConfig) {
-    if let Some(state) = STATE.get() {
+    let Some(state) = STATE.get() else {
+        return;
+    };
+    let resumed = {
         let mut guard = state.write();
+        let was_paused = matches!(guard.policy, Policy::Paused { .. });
         guard.cfg = cfg;
         guard.policy = decide(&guard.signals, &guard.cfg);
+        was_paused && !matches!(guard.policy, Policy::Paused { .. })
+    };
+    if resumed {
+        resume_notify().notify_one();
     }
 }
 
@@ -311,6 +349,12 @@ pub fn set_signed_out(signed_out: bool) {
     let prev = SIGNED_OUT.swap(signed_out, Ordering::AcqRel);
     if prev != signed_out {
         log::info!("[scheduler_gate] signed_out {} -> {}", prev, signed_out);
+        // #2831: signing back in (true -> false) is a transition out of
+        // `Policy::Paused { SignedOut }`. Wake any periodic loop so background
+        // sync restarts immediately rather than at the next tick boundary.
+        if prev && !signed_out {
+            resume_notify().notify_one();
+        }
     }
 }
 
@@ -325,6 +369,11 @@ pub fn set_signed_out(signed_out: bool) {
     let prev = test_state::set_signed_out_for(id, signed_out);
     if prev != signed_out {
         log::info!("[scheduler_gate] signed_out {} -> {}", prev, signed_out);
+        // #2831: mirror the production sign-in wake so tests exercise the
+        // same resume-notify path (true -> false fires the loop wake).
+        if prev && !signed_out {
+            resume_notify().notify_one();
+        }
     }
 }
 
@@ -700,4 +749,89 @@ mod tests {
         let p2 = try_acquire_llm_permit().expect("permit free after drop");
         drop(p2);
     }
+
+    /// #2831: both the firing side (`update_config` / `set_signed_out`) and the
+    /// waiting side (the periodic loop) must observe the *same* `Notify`, so
+    /// `resume_notify` must hand back one process-wide instance.
+    #[tokio::test]
+    async fn resume_notify_is_a_stable_singleton() {
+        let _g = lock();
+        assert!(
+            Arc::ptr_eq(&resume_notify(), &resume_notify()),
+            "resume_notify must return one shared instance"
+        );
+    }
+
+    /// A `notify_one()` wakes a task parked on `notified()` — the mechanism the
+    /// periodic loop relies on to resume early. Proves the singleton wiring
+    /// end-to-end (fire on one handle, wake on another).
+    #[tokio::test]
+    async fn resume_notify_wakes_a_parked_waiter() {
+        let _g = lock();
+        let waiter = resume_notify();
+        let parked = tokio::spawn(async move { waiter.notified().await });
+        // Yield so the spawned task reaches `.notified()` before we fire.
+        tokio::task::yield_now().await;
+        resume_notify().notify_one();
+        timeout(TokioDuration::from_secs(1), parked)
+            .await
+            .expect("parked waiter must wake promptly after notify_one")
+            .expect("waiter task must not panic");
+    }
+
+    /// #2831 wiring: a paused→running `update_config` transition and a
+    /// sign-in (`set_signed_out` true→false) each fire the resume notify.
+    ///
+    /// Seeds `STATE` directly (the test module can reach it) so `update_config`
+    /// / `set_signed_out` are live without spawning the real sampler. We drive
+    /// `Off → AlwaysOn`, which is deterministically `Paused → Aggressive`
+    /// regardless of any policy a prior test left behind, so the transition —
+    /// and thus the `notify_one()` — is guaranteed.
+    #[tokio::test]
+    async fn resume_transitions_fire_the_notify() {
+        use crate::openhuman::config::SchedulerGateMode;
+        let _g = lock();
+
+        // Ensure STATE is initialised. `set` is a no-op if an earlier test
+        // already promoted it — that's fine, we re-drive the transition below.
+        let cfg = SchedulerGateConfig {
+            mode: SchedulerGateMode::Off,
+            ..Default::default()
+        };
+        let signals = Signals::sample();
+        let policy = decide(&signals, &cfg);
+        let _ = STATE.set(Arc::new(RwLock::new(State {
+            cfg,
+            signals,
+            policy,
+        })));
+
+        // --- update_config: Paused -> running fires the notify ---
+        let waiter = resume_notify();
+        let parked = tokio::spawn(async move { waiter.notified().await });
+        tokio::task::yield_now().await;
+        update_config(SchedulerGateConfig {
+            mode: SchedulerGateMode::Off,
+            ..Default::default()
+        }); // -> Paused { UserDisabled }
+        update_config(SchedulerGateConfig {
+            mode: SchedulerGateMode::AlwaysOn,
+            ..Default::default()
+        }); // Paused -> Aggressive => resume fires
+        timeout(TokioDuration::from_secs(1), parked)
+            .await
+            .expect("update_config un-pause must wake the resume waiter")
+            .expect("waiter task must not panic");
+
+        // --- set_signed_out true -> false fires the notify ---
+        let waiter2 = resume_notify();
+        let parked2 = tokio::spawn(async move { waiter2.notified().await });
+        tokio::task::yield_now().await;
+        set_signed_out(true);
+        set_signed_out(false); // true -> false => resume fires
+        timeout(TokioDuration::from_secs(1), parked2)
+            .await
+            .expect("sign-in (signed_out true->false) must wake the resume waiter")
+            .expect("waiter task must not panic");
+    }
 }