diff --git a/better_memory/services/relevant.py b/better_memory/services/relevant.py
index 221b612..5ee7335 100644
--- a/better_memory/services/relevant.py
+++ b/better_memory/services/relevant.py
@@ -105,6 +105,7 @@ def _vec_qualifiers(
     id_col: str,
     query_vector: list[float] | None,
     vec_floor: float,
+    candidate_ids: list[str] | None = None,
 ) -> dict[str, int]:
     """id -> vec rank for rows within the cosine floor.
 
@@ -117,20 +118,37 @@ def _vec_qualifiers(
     L2 distance, NOT squared L2. The floor comparison below is against the
     plain (unsquared) distance only; the squared-distance branch that a
     defensive dual-check would need is dead code and has been omitted.
+
+    ``candidate_ids`` scopes the result to the caller's candidate set,
+    matching ``ReflectionService._vec_ranks`` (reflection.py) -- sqlite-vec
+    kNN accepts only ``embedding MATCH ? AND k = ?`` (no extra predicates),
+    so we fetch top-k and filter in Python. ``k`` scales with
+    ``max(len(candidate_ids), 50)`` so the window is dense over what the
+    caller actually ranks, instead of being drowned by neighbours from
+    other projects or retired records (whose vectors persist -- retire
+    only UPDATEs status). ``None`` preserves the global-scope behaviour
+    for callers (e.g. ``SqliteBackend.relevance_ranks``) that intentionally
+    want the unfiltered rank map; ``[]`` short-circuits to ``{}``.
     """
     if conn is None or query_vector is None:
         return {}
+    if candidate_ids is not None and not candidate_ids:
+        return {}
     max_dist = math.sqrt(2.0 * (1.0 - vec_floor))
+    k = max(len(candidate_ids), 50) if candidate_ids is not None else 50
     try:
         rows = conn.execute(
             f"SELECT {id_col}, distance FROM {table} "
             f"WHERE embedding MATCH ? AND k = ? ORDER BY distance",
-            (sqlite_vec.serialize_float32(query_vector), 50),
+            (sqlite_vec.serialize_float32(query_vector), k),
         ).fetchall()
     except sqlite3.OperationalError:
         return {}
+    wanted: set[str] | None = set(candidate_ids) if candidate_ids is not None else None
     out: dict[str, int] = {}
     for row in rows:
+        if wanted is not None and row[0] not in wanted:
+            continue
         if float(row[1]) <= max_dist:
             out[row[0]] = len(out)
     return out
@@ -213,8 +231,26 @@ def retrieve_relevant(
 
     bm = _bm25_qualifiers(conn, query)
     qvec = sync_embedder.embed_text(query) if sync_embedder is not None else None
-    vec_r = _vec_qualifiers(conn, "reflection_embeddings", "reflection_id", qvec, vec_floor)
-    vec_s = _vec_qualifiers(conn, "semantic_embeddings", "memory_id", qvec, vec_floor)
+    # Collect the actual candidate id set before the vec query so kNN is
+    # restricted to it (issue #104): the vec tables are global across
+    # projects and retained/retired rows, so an unbounded k=50 kNN can be
+    # entirely filled by neighbours the caller will never rank -- starving
+    # the evidence gate for the caller's own reflections/semantics.
+    refl_bucket_order = ["do", "dont"] + (["neutral"] if include_neutral else [])
+    refl_candidate_ids = [
+        str(r.get("id"))
+        for bucket in refl_bucket_order
+        for r in (buckets.get(bucket) or [])
+    ]
+    sem_candidate_ids = [str(getattr(s, "id", "")) for s in (semantic or [])]
+    vec_r = _vec_qualifiers(
+        conn, "reflection_embeddings", "reflection_id", qvec, vec_floor,
+        refl_candidate_ids,
+    )
+    vec_s = _vec_qualifiers(
+        conn, "semantic_embeddings", "memory_id", qvec, vec_floor,
+        sem_candidate_ids,
+    )
     keywords = extract_keywords(query)     # fallback evidence only
 
     fts_unavailable = conn is None
@@ -255,8 +291,7 @@ def retrieve_relevant(
     rank_map: dict[tuple[str, str], int] = raw_rank_map or {}
 
     refl_candidates: list[dict] = []
-    order = ["do", "dont"] + (["neutral"] if include_neutral else [])
-    for bucket in order:
+    for bucket in refl_bucket_order:
         for r in buckets.get(bucket, []) or []:
             r_id = str(r.get("id"))
             title = str(r.get("title") or "")
diff --git a/tests/services/test_relevant.py b/tests/services/test_relevant.py
index d7195a2..c22a17c 100644
--- a/tests/services/test_relevant.py
+++ b/tests/services/test_relevant.py
@@ -144,6 +144,33 @@ def test_vec_below_floor_does_not_qualify(self, conn):
         )
         assert out == []
 
+    def test_orphan_vector_row_does_not_leak_via_vec_gate(self, conn):
+        # Issue #104: the vec table is global and its rows persist across
+        # retire (retire only UPDATEs status; the vector stays). An
+        # "orphan" vector -- one whose metadata row was retired or deleted
+        # so backend.retrieve does not surface it as a candidate -- must
+        # not qualify anything through the vec leg. Prior to the fix,
+        # _vec_qualifiers issued an unfiltered kNN, so the orphan showed
+        # up in vec_r but happened to be harmless only because the RRF
+        # loop keys on the caller's candidate list -- this test locks in
+        # that the candidate scoping is now enforced at the vec-query
+        # layer too, so orphan ids never appear in vec_r at all.
+        from better_memory.services.relevant import _vec_qualifiers
+        emb = DirectedEmbedder("orphan trigger")
+        # Insert a vec row for a reflection id that has NO metadata row.
+        _embed_reflection(conn, "r-orphan", emb._vec("orphan trigger"))
+        qvec = emb._vec("orphan trigger")
+        # candidate_ids empty -> short-circuit
+        assert _vec_qualifiers(
+            conn, "reflection_embeddings", "reflection_id",
+            qvec, 0.55, [],
+        ) == {}
+        # candidate_ids restricts what returns from the kNN scan
+        assert _vec_qualifiers(
+            conn, "reflection_embeddings", "reflection_id",
+            qvec, 0.55, ["r-different"],
+        ) == {}
+
 
 class TestWilsonRanking:
     def test_wilson_ranks_among_qualifiers(self, conn):