diff --git a/docs/forge.md b/docs/forge.md
index f3bf9d00..a2dfd793 100644
--- a/docs/forge.md
+++ b/docs/forge.md
@@ -70,9 +70,46 @@ Background task spawned once per harness boot. Polls
 `GET /api/v1/notifications?all=false` every 30 seconds (Forgejo's
 unread-only filter), formats each notification as a broker
 `Wake { from: "forge" }` message, and delivers it to the agent's own
-inbox so claude's normal turn loop picks it up. Mark-read happens
-after successful delivery so a failed-delivery notification
-resurfaces on the next tick.
+inbox so claude's normal turn loop picks it up.
+
+### Mark-read on read, not on delivery
+
+Delivered conversation threads are deliberately left **unread** in
+forge. The hive-forge read-before-comment guard keys off forge's own
+notification read-state (`GET /notifications?all=false`) to refuse a
+comment when a thread has unread activity by others — so the agent
+reading the thread via the CLI (`hive-forge comments` / `view`,
+which `PATCH`es `/notifications/threads/{id}`) is the single
+mark-read point. If `forge_notify` marked threads read on delivery,
+that unread signal would be consumed before the agent acts and the
+guard could never fire.
+
+Because a delivered thread stays unread, it reappears in every
+`?all=false` poll. An in-memory **delivery-dedupe cursor** (thread
+id → last-delivered `updated_at`, held in the poll loop) stops the
+same version from re-firing a wake; a new comment bumps `updated_at`
+so genuinely new activity re-delivers. The cursor is pure anti-spam,
+not a correctness oracle: lost on harness restart it just
+re-delivers currently-unread threads once (harmless — `recv`
+tolerates redelivery), so it carries none of the persisted-mirror
+fragility that ruled out an on-disk seen-cursor. Each poll prunes
+the cursor to the threads still in the unread set. A failed wake
+delivery is left unread **and** out of the cursor, so it resurfaces
+next tick.
+
+Two paths still mark-read directly (no read-before-comment value):
+self-echo notifications (the agent's own writes, see below) and
+`HIVE_FORGE_NOTIFY_SKIP_REASONS` drop-listed reasons.
+
+> Note: the unread list grows for threads the agent never reads via
+> the CLI, since nothing else trims it. This does not affect guard
+> correctness (the guard does a per-thread, repo-scoped query) nor
+> wake delivery (Forgejo orders unread newest-first, so new activity
+> always lands in the polled window). Bounding the unread list via a
+> reason-independent firehose-reduction is a separate follow-up — the
+> existing auto-unsubscribe below is gated on a `reason` field that
+> this Forgejo's notification API does not actually emit, so it never
+> fires today.
 
 ### Activation gates (graceful no-ops)
 
diff --git a/hive-ag3nt/src/forge_notify.rs b/hive-ag3nt/src/forge_notify.rs
index 86e0753b..b99026a1 100644
--- a/hive-ag3nt/src/forge_notify.rs
+++ b/hive-ag3nt/src/forge_notify.rs
@@ -1,7 +1,13 @@
 //! Background Forgejo notification poller. Polls
 //! `GET /notifications?all=false` every 30s, formats each unread
 //! notification as a broker `Wake { from: "forge" }` message, and
-//! marks it read after delivery so failures resurface next tick.
+//! delivers it to the agent's inbox. Delivered threads are deliberately
+//! left UNREAD in forge — the hive-forge read-before-comment guard keys
+//! off forge's own unread-state, and the agent reading the thread via the
+//! CLI is what marks it read. An in-memory delivery-dedupe cursor
+//! (thread id → last-delivered `updated_at`) stops the still-unread
+//! notification from re-firing a wake every poll; self-echo and
+//! drop-listed notifications are still marked read directly.
 //!
 //! Activation gates, self-notification filtering, body excerpt +
 //! truncation + heading escape, wrapper formats (comment / review /
@@ -9,7 +15,7 @@
 //! reason drop-list, and auto-unsubscribe on broad watches all live
 //! in [`docs/forge.md::Notification poller`](../../../docs/forge.md).
 
-use std::collections::HashSet;
+use std::collections::{HashMap, HashSet};
 use std::fmt::Write as _;
 use std::path::{Path, PathBuf};
 use std::time::Duration;
@@ -139,6 +145,19 @@ pub async fn run(socket: PathBuf) {
     // across polls so we don't hammer DELETE on every cycle.
     let mut unsubbed_repos: HashSet<String> = HashSet::new();
 
+    // Delivery-dedupe cursor: notification thread id -> the `updated_at`
+    // of the version we last woke the agent for. We no longer mark a
+    // thread read on delivery (that would consume the unread signal the
+    // hive-forge read-before-comment guard relies on), so this in-memory
+    // map is what stops the same unread notification from re-firing a
+    // wake every poll. A new comment bumps `updated_at`, so the thread
+    // re-delivers. This is purely anti-spam, NOT a correctness oracle:
+    // lost on harness restart it just re-delivers currently-unread
+    // threads once (harmless — recv tolerates redelivery), so it carries
+    // none of the persisted-mirror fragility that sank the on-disk
+    // cursor approach.
+    let mut delivered: HashMap<u64, String> = HashMap::new();
+
     loop {
         interval.tick().await;
         poll_once(
@@ -148,6 +167,7 @@ pub async fn run(socket: PathBuf) {
             &socket,
             keep_subscriptions,
             &mut unsubbed_repos,
+            &mut delivered,
             &own_login,
             &skip_reasons,
         )
@@ -723,9 +743,9 @@ fn days_from_civil(y: i64, m: i64, d: i64) -> i64 {
 
 #[allow(
     clippy::too_many_arguments,
-    reason = "the notification poll's config + mutable subscription state, \
-              wired once from the poll loop; a struct would just move the \
-              same fields one level out"
+    reason = "the notification poll's config + mutable subscription / \
+              delivery-dedupe state, wired once from the poll loop; a struct \
+              would just move the same fields one level out"
 )]
 #[allow(
     clippy::too_many_lines,
@@ -740,6 +760,7 @@ async fn poll_once(
     socket: &Path,
     keep_subscriptions: bool,
     unsubbed_repos: &mut HashSet<String>,
+    delivered: &mut HashMap<u64, String>,
     own_login: &str,
     skip_reasons: &[String],
 ) {
@@ -784,6 +805,17 @@ async fn poll_once(
             continue;
         };
 
+        // Delivery-dedupe: we no longer mark threads read on delivery, so
+        // an unread notification reappears in every `?all=false` poll.
+        // Skip it silently unless its `updated_at` advanced since the
+        // version we last delivered a wake for (i.e. genuinely new
+        // activity). See the `delivered` cursor note in `run`.
+        let updated_at = notif["updated_at"].as_str().unwrap_or("").to_owned();
+        if !should_deliver(delivered, id, &updated_at) {
+            debug!(%id, "forge_notify: skipping (already delivered this version)");
+            continue;
+        }
+
         // Reason drop-list: suppress noisy reasons; null/unknown pass
         // through so directed signals stay deliverable (see
         // `docs/forge.md::Reason drop-list`).
@@ -809,10 +841,10 @@ async fn poll_once(
             body,
             transient: false,
         };
-        let delivered = crate::client::request::<_, hive_sh4re::Response>(socket, &req)
+        let deliver_result = crate::client::request::<_, hive_sh4re::Response>(socket, &req)
             .await
             .map(|_| ());
-        match delivered {
+        match deliver_result {
             Ok(()) => {
                 debug!(%id, "forge_notify: delivered");
             }
@@ -822,9 +854,13 @@ async fn poll_once(
             }
         }
 
-        // Mark as read only after successful delivery so a failed-delivery
-        // notification resurfaces on the next poll tick.
-        mark_read(client, forge_url, token, id).await;
+        // Record the delivered version in the dedupe cursor INSTEAD of
+        // marking the thread read. Leaving it unread is deliberate: the
+        // hive-forge read-before-comment guard keys off forge's own
+        // unread-state, and the agent reading the thread via the CLI is
+        // what marks it read. A failed delivery (above) is left
+        // unrecorded so it re-delivers next tick.
+        delivered.insert(id, updated_at);
 
         // Auto-unsubscribe from broad repo watches after delivering a
         // `subscribed` notification. Gated by HIVE_FORGE_KEEP_SUBSCRIPTIONS
@@ -856,12 +892,34 @@ async fn poll_once(
             }
         }
     }
+
+    // Prune the dedupe cursor down to the threads still present in this
+    // poll's unread set. Once the agent reads a thread (marking it read
+    // via the CLI) it drops out of `?all=false`, so its cursor entry is
+    // dead weight; dropping it bounds the map to the current unread size.
+    // If such a thread later goes unread again it carries a fresh
+    // `updated_at` and re-delivers correctly.
+    let current_ids: HashSet<u64> = notifications
+        .iter()
+        .filter_map(|n| n["id"].as_u64())
+        .collect();
+    delivered.retain(|id, _| current_ids.contains(id));
+}
+
+/// Whether a notification should be delivered as a wake given the
+/// delivery-dedupe cursor. Delivers when the thread has never been
+/// delivered, or when its `updated_at` advanced since the last delivered
+/// version (genuinely new activity). Pure for unit testing.
+fn should_deliver(delivered: &HashMap<u64, String>, id: u64, updated_at: &str) -> bool {
+    delivered.get(&id).is_none_or(|seen| seen != updated_at)
 }
 
 /// Mark a notification thread as read. Best-effort — logs on failure but
-/// does not abort the poll loop. A notification left unread will resurface
-/// on the next poll tick (desirable for delivery failures; for self-echo
-/// silencing we call this without prior delivery).
+/// does not abort the poll loop. Called only on the self-echo and
+/// drop-listed paths (the agent's own writes / explicitly-suppressed
+/// reasons) — delivered threads are deliberately left unread for the
+/// read-before-comment guard, and a failed delivery is left unread + out
+/// of the dedupe cursor so it resurfaces on the next poll tick.
 async fn mark_read(client: &reqwest::Client, forge_url: &str, token: &str, id: u64) {
     let mark_url = format!("{forge_url}/api/v1/notifications/threads/{id}");
     match client
@@ -886,6 +944,38 @@ async fn mark_read(client: &reqwest::Client, forge_url: &str, token: &str, id: u
 mod tests {
     use super::*;
 
+    #[test]
+    fn should_deliver_when_thread_never_seen() {
+        let delivered = HashMap::new();
+        assert!(should_deliver(&delivered, 42, "2026-06-22T16:00:00Z"));
+    }
+
+    #[test]
+    fn should_not_deliver_same_version_again() {
+        // The dedupe case: an unread thread reappears every poll with the
+        // same `updated_at` — must not re-fire a wake.
+        let mut delivered = HashMap::new();
+        delivered.insert(42, "2026-06-22T16:00:00Z".to_owned());
+        assert!(!should_deliver(&delivered, 42, "2026-06-22T16:00:00Z"));
+    }
+
+    #[test]
+    fn should_deliver_when_updated_at_advanced() {
+        // A new comment bumps `updated_at` → genuinely new activity →
+        // deliver again.
+        let mut delivered = HashMap::new();
+        delivered.insert(42, "2026-06-22T16:00:00Z".to_owned());
+        assert!(should_deliver(&delivered, 42, "2026-06-22T16:05:00Z"));
+    }
+
+    #[test]
+    fn should_deliver_tracks_per_thread() {
+        // A cursor for one thread says nothing about another.
+        let mut delivered = HashMap::new();
+        delivered.insert(42, "2026-06-22T16:00:00Z".to_owned());
+        assert!(should_deliver(&delivered, 99, "2026-06-22T16:00:00Z"));
+    }
+
     #[test]
     fn escape_md_headings_escapes_top_level_atx() {
         // Argus reviews start with `## argus review`, which would