diff --git a/engram/dist/engram b/engram/dist/engram
index 2236bee..1d7defa 100755
Binary files a/engram/dist/engram and b/engram/dist/engram differ
diff --git a/engram/dist/engram.c b/engram/dist/engram.c
index 9c3ffce..00aa3cc 100644
--- a/engram/dist/engram.c
+++ b/engram/dist/engram.c
@@ -24,6 +24,7 @@ el_val_t route_forget(el_val_t method, el_val_t path, el_val_t body);
 el_val_t route_save(el_val_t method, el_val_t path, el_val_t body);
 el_val_t route_load(el_val_t method, el_val_t path, el_val_t body);
 el_val_t route_health(el_val_t method, el_val_t path, el_val_t body);
+el_val_t route_embed_backfill(el_val_t method, el_val_t path, el_val_t body);
 el_val_t route_sync(el_val_t method, el_val_t path, el_val_t body);
 el_val_t route_load_merge(el_val_t method, el_val_t path, el_val_t body);
 el_val_t route_emit_ise(el_val_t method, el_val_t path, el_val_t body);
@@ -271,6 +272,17 @@ el_val_t route_health(el_val_t method, el_val_t path, el_val_t body) {
   return 0;
 }
 
+el_val_t route_embed_backfill(el_val_t method, el_val_t path, el_val_t body) {
+  el_val_t n = query_int(path, EL_STR("n"), 32);
+  el_val_t result = engram_embed_backfill(n);
+  el_val_t done = json_get_float(result, EL_STR("embedded"));
+  if (done > el_from_float(0.0)) {
+    el_val_t saved = persist_canonical();
+  }
+  return result;
+  return 0;
+}
+
 el_val_t route_sync(el_val_t method, el_val_t path, el_val_t body) {
   el_val_t dir_raw = env(EL_STR("ENGRAM_DATA_DIR"));
   el_val_t dir = ({ el_val_t _if_result_20 = 0; if (str_eq(dir_raw, EL_STR(""))) { _if_result_20 = (EL_STR("/tmp/engram")); } else { _if_result_20 = (dir_raw); } _if_result_20; });
@@ -432,6 +444,9 @@ el_val_t handle_request(el_val_t method, el_val_t path, el_val_t body) {
   if (str_eq(method, EL_STR("GET")) && str_eq(clean, EL_STR("/api/sync"))) {
     return route_sync(method, path, body);
   }
+  if (str_eq(clean, EL_STR("/api/embed-backfill"))) {
+    return route_embed_backfill(method, path, body);
+  }
   return el_str_concat(el_str_concat(EL_STR("{\"error\":\"not found\",\"path\":\""), clean), EL_STR("\"}"));
   return 0;
 }
diff --git a/engram/src/server.el b/engram/src/server.el
index 5e72136..6689ed8 100644
--- a/engram/src/server.el
+++ b/engram/src/server.el
@@ -246,6 +246,28 @@ fn route_health(method: String, path: String, body: String) -> String {
     "{\"status\":\"ok\",\"engine\":\"engram-runtime-native\"}"
 }
 
+// route_embed_backfill — GET/POST /api/embed-backfill?n=48
+//
+// (2026-07-25 self-review) The lazy embedding backfill runs only inside
+// engram_activate, and nothing in production calls /api/activate on this
+// store — the soul's curiosity loop activates its own in-process graph.
+// After a restart from a snapshot without vectors, embedded_count stalled
+// at 93/12175 and would never recover. This route lets the soul's
+// heartbeat pump the backfill explicitly (48/min clears a 12k backlog in
+// ~4h). Persists the canonical snapshot whenever new vectors were
+// generated — the 2026-07-25 regression happened precisely because 3747
+// in-RAM embeddings were never snapshotted before a restart. Self-
+// limiting: once coverage is full, embedded=0 and no save occurs.
+fn route_embed_backfill(method: String, path: String, body: String) -> String {
+    let n: Int = query_int(path, "n", 32)
+    let result: String = engram_embed_backfill(n)
+    let done: Float = json_get_float(result, "embedded")
+    if done > 0.0 {
+        let saved: Int = persist_canonical()
+    }
+    return result
+}
+
 // route_sync — return a snapshot of non-ISE/non-Working nodes for the soul daemon
 // to merge into its in-process graph via engram_load_merge.
 //
@@ -495,6 +517,11 @@ fn handle_request(method: String, path: String, body: String) -> String {
         return route_sync(method, path, body)
     }
 
+    // Embedding backfill — pumped by the soul heartbeat (2026-07-25)
+    if str_eq(clean, "/api/embed-backfill") {
+        return route_embed_backfill(method, path, body)
+    }
+
     "{\"error\":\"not found\",\"path\":\"" + clean + "\"}"
 }
 
diff --git a/lang/releases/v1.0.0-20260501/el_runtime.c b/lang/releases/v1.0.0-20260501/el_runtime.c
index cf31d3c..04f0948 100644
--- a/lang/releases/v1.0.0-20260501/el_runtime.c
+++ b/lang/releases/v1.0.0-20260501/el_runtime.c
@@ -5680,6 +5680,24 @@ void el_cgi_init(el_val_t name, el_val_t dharma_id, el_val_t principal,
 #define ENGRAM_BLL_TAU  (-3.0)
 #define ENGRAM_BLL_S    0.4
 
+/* Short-term inhibition-of-return (2026-07-25 self-review).
+ * Lebiere & Best 2009 ("Balancing Long-Term Reinforcement and Short-Term
+ * Inhibition", CogSci): subtract ln(1 + (t_n/t_s)^-d_s) from activation,
+ * where t_n = time since the MOST RECENT access. With their best-fit
+ * d_s = 1.0 the exp of the subtraction reduces to the clean multiplier
+ *     m(t_n) = t_n / (t_n + t_s)
+ * applied to raw_wm in Pass 2. Immediately after a promotion the node's
+ * score is crushed (m → 0), then self-heals as a power law — producing an
+ * emergent round-robin over WM candidates instead of winner-take-all.
+ * This replaces the non-decaying suppression_count as the damping
+ * mechanism (that counter never entered the score at all — it only ever
+ * pushed nodes TOWARD surfacing via breakthrough; kept for that role).
+ * t_s = 4× the ~30 s curiosity-scan interval per the paper's guidance
+ * (t_s ≈ peak re-occurrence lag). A node promoted every scan holds
+ * m ≈ 0.2 until it loses its slot; after ~2 min unretrieved, m ≥ 0.5.
+ * Source: act-r.psy.cmu.edu/.../894Cogsci09-Lebiere-Best.pdf */
+#define ENGRAM_STI_TS   120.0
+
 /* qsort comparator — descending double, used by WM cap enforcement. */
 static int engram_cmp_double_desc(const void* a, const void* b) {
     double da = *(const double*)a;
@@ -7830,6 +7848,23 @@ el_val_t engram_activate(el_val_t query, el_val_t depth) {
             raw_wm += ENGRAM_EMBED_WM_WEIGHT
                     * (cosq[i] - ENGRAM_EMBED_S0) / (1.0 - ENGRAM_EMBED_S0);
         }
+        /* Short-term inhibition-of-return (2026-07-25, Lebiere-Best):
+         * damp by t_n/(t_n + t_s) where t_n = seconds since the most
+         * recent recorded access (WM promotion / strengthen). Applied
+         * after the cosine term so the WHOLE score rotates — a node that
+         * just held a WM slot yields it even to semantically weaker
+         * competitors, and recovers as t_n grows. access_ts is recorded
+         * at promotion, so persistent WM residents self-inhibit. Nodes
+         * with no access history (never promoted) are uninhibited.
+         * Layer-0 override in Pass 3 still floors safety nodes. */
+        if (n->access_filled > 0) {
+            int32_t sti_last = (n->access_head + ENGRAM_BLL_K - 1)
+                             % ENGRAM_BLL_K;
+            double sti_tn = (double)(now_ms - n->access_ts[sti_last])
+                          / 1000.0;
+            if (sti_tn < 0.1) sti_tn = 0.1;
+            raw_wm *= sti_tn / (sti_tn + ENGRAM_STI_TS);
+        }
         /* Threshold gate: must exceed per-type threshold to enter working
          * memory. Type threshold replaces the old flat 0.2 filter. */
         if (raw_wm >= type_threshold) {
@@ -9080,6 +9115,43 @@ el_val_t engram_cosine_sim(el_val_t id_a, el_val_t id_b) {
     return el_from_float(eg_cosine(a->emb, b->emb, a->emb_dim));
 }
 
+/* engram_embed_backfill — explicitly drive the lazy embedding backfill.
+ * (2026-07-25 self-review.) The per-activate backfill (8 nodes/call) only
+ * runs inside engram_activate, and on the authoritative HTTP store nothing
+ * calls /api/activate in production — so embedded_count stalled at ~93 of
+ * 12k after a restart from a snapshot without vectors. This entry point
+ * lets the server expose a route the soul's heartbeat can pump. Embeds up
+ * to `count` eligible un-embedded nodes (clamped to [1, 64]), newest
+ * first, respecting the same eligibility rules and circuit breaker as the
+ * lazy path. Returns JSON: {"embedded":k,"embedded_count":m,"node_count":c}.
+ * k = 0 with the breaker open or embedder down — never blocks past the
+ * per-fetch timeout chain. */
+el_val_t engram_embed_backfill(el_val_t count) {
+    EngramStore* g = engram_get();
+    int64_t want = (int64_t)EL_INT(count);
+    if (want < 1)  want = 1;
+    if (want > 64) want = 64;
+    int64_t done = 0;
+    for (int64_t i = g->node_count - 1; i >= 0 && done < want; i--) {
+        EngramNode* n = &g->nodes[i];
+        if (n->emb || !eg_embed_eligible(n)) continue;
+        int32_t d = 0;
+        float* v = eg_embed_fetch(n->content, &d);
+        if (!v) break; /* embedder down / breaker open — stop this call */
+        n->emb = v; n->emb_dim = d;
+        done++;
+    }
+    int64_t total = 0;
+    for (int64_t i = 0; i < g->node_count; i++) {
+        if (g->nodes[i].emb) total++;
+    }
+    char buf[160];
+    snprintf(buf, sizeof(buf),
+             "{\"embedded\":%lld,\"embedded_count\":%lld,\"node_count\":%lld}",
+             (long long)done, (long long)total, (long long)g->node_count);
+    return el_wrap_str(el_strdup(buf));
+}
+
 /* engram_list_layers_json — serialized counterpart of engram_list_layers.
  * Returns a JSON array, sorted by activation_priority ascending. */
 el_val_t engram_list_layers_json(void) {
diff --git a/lang/releases/v1.0.0-20260501/el_runtime.h b/lang/releases/v1.0.0-20260501/el_runtime.h
index 87ac86b..a35d1ae 100644
--- a/lang/releases/v1.0.0-20260501/el_runtime.h
+++ b/lang/releases/v1.0.0-20260501/el_runtime.h
@@ -618,6 +618,7 @@ el_val_t  engram_neighbors_json(el_val_t node_id, el_val_t max_depth, el_val_t d
 el_val_t  engram_activate_json(el_val_t query, el_val_t depth);
 el_val_t  engram_stats_json(void);
 el_val_t  engram_cosine_sim(el_val_t id_a, el_val_t id_b);
+el_val_t  engram_embed_backfill(el_val_t count);
 el_val_t  engram_list_layers_json(void);
 /* Working memory introspection — count, mean weight, and top-N snapshot.
  * Ported from el-compiler/runtime on 2026-06-30 self-review. */