diff --git a/src/database/knowledge_dao.rs b/src/database/knowledge_dao.rs
index ffc368f..091cd9a 100644
--- a/src/database/knowledge_dao.rs
+++ b/src/database/knowledge_dao.rs
@@ -601,12 +601,41 @@ impl KnowledgeDao for SqliteKnowledgeDao {
         trace_db_call(cx, "delete", "delete_entity", |_span| {
             use schema::entities::dsl::*;
             let mut conn = self.connection.lock().expect("KnowledgeDao lock");
-            diesel::delete(entities.filter(id.eq(entity_id)))
-                .execute(conn.deref_mut())
-                .map(|_| ())
-                .map_err(|e| anyhow::anyhow!("Delete error: {}", e))
+
+            // entity_facts has a CHECK constraint requiring
+            // `object_entity_id IS NOT NULL OR object_value IS NOT NULL`.
+            // The FK on object_entity_id is ON DELETE SET NULL — but
+            // facts that pointed at the deleted entity *only* via the
+            // entity reference (the common case for relational facts
+            // like "Alice is_friend_of Bob") have no object_value, so
+            // SET NULL would leave them with both NULLs and the CHECK
+            // aborts the whole DELETE. Pre-delete those facts in a
+            // transaction so the CASCADE / SET NULL chain on what
+            // remains can fire cleanly.
+            //
+            // Long-term fix is to change the FK to ON DELETE CASCADE
+            // via a table-rebuild migration, but the DAO-side workaround
+            // is sufficient and less invasive.
+            conn.transaction::<(), diesel::result::Error, _>(|conn| {
+                use schema::entity_facts::dsl as ef;
+                diesel::delete(
+                    ef::entity_facts
+                        .filter(ef::object_entity_id.eq(entity_id))
+                        .filter(ef::object_value.is_null()),
+                )
+                .execute(conn)?;
+
+                diesel::delete(entities.filter(id.eq(entity_id))).execute(conn)?;
+                Ok(())
+            })
+            .map_err(|e| anyhow::anyhow!("Delete error: {}", e))
+        })
+        .map_err(|e| {
+            // Surface the actual diesel error string before collapsing
+            // to the opaque DbErrorKind::QueryError.
+            log::warn!("delete_entity({}) failed: {}", entity_id, e);
+            DbError::new(DbErrorKind::QueryError)
         })
-        .map_err(|_| DbError::new(DbErrorKind::QueryError))
     }
 
     fn merge_entities(
@@ -1334,6 +1363,77 @@ mod tests {
         );
     }
 
+    #[test]
+    fn delete_entity_clears_relational_facts_that_would_violate_check() {
+        // entity_facts has a CHECK that at least one of object_entity_id /
+        // object_value is non-null. The FK on object_entity_id is
+        // ON DELETE SET NULL, which would leave purely-relational facts
+        // (subject + predicate + object_entity_id, no object_value)
+        // with both nulls and abort the delete. The DAO pre-deletes
+        // those rows in a transaction so the parent delete can succeed.
+        let cx = opentelemetry::Context::new();
+        let conn = connection_with_fks_on();
+        let alice = create_user(&conn, "alice");
+        create_persona_row(&conn, alice, "default");
+
+        let mut dao = SqliteKnowledgeDao::from_connection(conn.clone());
+        let bob = make_entity(&mut dao, "Bob");
+        let carol = make_entity(&mut dao, "Carol");
+
+        // A relational fact where Carol is the object — exactly the
+        // shape the CHECK + SET NULL combination would otherwise break.
+        let (rel_fact, _) = dao
+            .upsert_fact(
+                &cx,
+                InsertEntityFact {
+                    subject_entity_id: bob.id,
+                    predicate: "is_friend_of".to_string(),
+                    object_entity_id: Some(carol.id),
+                    object_value: None,
+                    source_photo: None,
+                    source_insight_id: None,
+                    confidence: 0.6,
+                    status: "active".to_string(),
+                    created_at: 0,
+                    persona_id: "default".to_string(),
+                    user_id: alice,
+                },
+            )
+            .unwrap();
+
+        // A typed fact where Bob is the subject — should survive.
+        add_fact(&mut dao, bob.id, "has_age", "30", alice, "default");
+
+        // Delete Carol — should succeed (relational fact pre-deleted).
+        dao.delete_entity(&cx, carol.id).unwrap();
+
+        assert!(
+            dao.get_entity_by_id(&cx, carol.id).unwrap().is_none(),
+            "Carol should be deleted"
+        );
+        // The relational fact about Carol should be gone (pre-deleted by
+        // the DAO's transaction, not SET NULL'd).
+        let bob_facts = dao
+            .get_facts_for_entity(
+                &cx,
+                bob.id,
+                &PersonaFilter::Single {
+                    user_id: alice,
+                    persona_id: "default".to_string(),
+                },
+            )
+            .unwrap();
+        assert!(
+            !bob_facts.iter().any(|f| f.id == rel_fact.id),
+            "relational fact pointing at Carol should be removed"
+        );
+        // The typed fact survives.
+        assert!(
+            bob_facts.iter().any(|f| f.predicate == "has_age"),
+            "typed fact about Bob should survive Carol's deletion"
+        );
+    }
+
     #[test]
     fn upsert_entity_collapses_near_duplicate_by_embedding() {
         // The agent's pre-flight check uses FTS5 prefix tokens, which