diff --git a/src/openhuman/memory_store/vectors/store.rs b/src/openhuman/memory_store/vectors/store.rs
index 90fd8594d..5dae581a7 100644
--- a/src/openhuman/memory_store/vectors/store.rs
+++ b/src/openhuman/memory_store/vectors/store.rs
@@ -326,18 +326,32 @@ impl VectorStore {
             })?
             .collect::<rusqlite::Result<Vec<_>>>()?;
 
-        let mut scored: Vec<SearchResult> = rows
+        let scanned = rows.len();
+
+        // Score-only intermediate: keep metadata as the raw JSON string instead
+        // of parsing it here. We scan every vector in the namespace but return
+        // only `limit` rows, so parsing the metadata of all N candidates would
+        // throw away all but `limit` parses. Defer the parse until after the
+        // truncation below, where it runs `limit` times instead of N.
+        struct ScoredRow {
+            score: f64,
+            id: String,
+            namespace: String,
+            text: String,
+            meta_str: String,
+        }
+
+        let mut scored: Vec<ScoredRow> = rows
             .into_iter()
-            .map(|(id, ns, text, blob, meta_str)| {
+            .map(|(id, namespace, text, blob, meta_str)| {
                 let stored_vec = bytes_to_vec(&blob);
                 let score = cosine_similarity(query_vec, &stored_vec);
-                let metadata = serde_json::from_str(&meta_str).unwrap_or(serde_json::Value::Null);
-                SearchResult {
-                    id,
-                    namespace: ns,
-                    text,
+                ScoredRow {
                     score,
-                    metadata,
+                    id,
+                    namespace,
+                    text,
+                    meta_str,
                 }
             })
             .collect();
@@ -350,14 +364,37 @@ impl VectorStore {
         });
         scored.truncate(limit);
 
+        // Parse metadata only for the rows that survived truncation. Invalid
+        // JSON falls back to Null, but log it so data issues stay diagnosable.
+        let results: Vec<SearchResult> = scored
+            .into_iter()
+            .map(|row| {
+                let metadata = serde_json::from_str(&row.meta_str).unwrap_or_else(|err| {
+                    tracing::debug!(
+                        target: "embeddings.store",
+                        "[vector-store] invalid metadata json: id={}, ns={}, err={err}",
+                        row.id,
+                        row.namespace,
+                    );
+                    serde_json::Value::Null
+                });
+                SearchResult {
+                    id: row.id,
+                    namespace: row.namespace,
+                    text: row.text,
+                    score: row.score,
+                    metadata,
+                }
+            })
+            .collect();
+
         tracing::trace!(
             target: "embeddings.store",
-            "[vector-store] search_by_vector: ns={namespace}, scanned={}, returned={}",
-            scored.len() + scored.capacity() - scored.len(), // approximate total before truncate
-            scored.len()
+            "[vector-store] search_by_vector: ns={namespace}, scanned={scanned}, returned={}",
+            results.len()
         );
 
-        Ok(scored)
+        Ok(results)
     }
 
     // ── Delete / management ──────────────────────────────────
diff --git a/src/openhuman/memory_store/vectors/store_tests.rs b/src/openhuman/memory_store/vectors/store_tests.rs
index c5b9e9af6..2d3c656d5 100644
--- a/src/openhuman/memory_store/vectors/store_tests.rs
+++ b/src/openhuman/memory_store/vectors/store_tests.rs
@@ -295,6 +295,78 @@ fn search_by_vector_limit_zero() {
         .is_empty());
 }
 
+/// Metadata is parsed only for rows that survive truncation, so the parse
+/// must align with the post-sort order — each returned row must carry its
+/// own metadata, and dropped rows must not appear. This pins the deferred
+/// parse against an off-by-one or mis-zipped mapping after sort/truncate.
+#[test]
+fn search_by_vector_returns_metadata_of_surviving_rows() {
+    let store = fake_store(3);
+    store
+        .insert_with_vector("near", "ns", "t", &[1.0, 0.0, 0.0], json!({"tag": "near"}))
+        .unwrap();
+    store
+        .insert_with_vector("mid", "ns", "t", &[0.7, 0.7, 0.0], json!({"tag": "mid"}))
+        .unwrap();
+    store
+        .insert_with_vector("far", "ns", "t", &[0.0, 0.0, 1.0], json!({"tag": "far"}))
+        .unwrap();
+
+    let results = store.search_by_vector("ns", &[1.0, 0.0, 0.0], 2).unwrap();
+
+    assert_eq!(results.len(), 2, "limit should drop the least similar row");
+    assert_eq!(results[0].id, "near");
+    assert_eq!(results[1].id, "mid");
+    assert!(
+        results.iter().all(|hit| hit.id != "far"),
+        "the truncated row must not leak into the results"
+    );
+    // The deferred parse must attach each row's own metadata, not a neighbour's.
+    for hit in &results {
+        assert_eq!(
+            hit.metadata.get("tag").and_then(|v| v.as_str()),
+            Some(hit.id.as_str()),
+            "metadata tag should match the row id for {}",
+            hit.id
+        );
+    }
+}
+
+/// A row with corrupt metadata JSON (e.g. a hand-edited or partially written
+/// DB) must not break the search — the deferred parse falls back to `Null`
+/// rather than dropping the row or erroring. Inserted raw because
+/// `insert_with_vector` always serializes valid JSON.
+#[test]
+fn search_by_vector_falls_back_to_null_on_invalid_metadata() {
+    let store = fake_store(3);
+    {
+        let conn = store.conn.lock();
+        conn.execute(
+            "INSERT INTO vectors (id, namespace, text, embedding, metadata, created_at, updated_at)
+             VALUES (?1, ?2, ?3, ?4, ?5, ?6, ?7)",
+            rusqlite::params![
+                "bad",
+                "ns",
+                "t",
+                vec_to_bytes(&[1.0, 0.0, 0.0]),
+                "{not valid json",
+                0.0_f64,
+                0.0_f64
+            ],
+        )
+        .unwrap();
+    }
+
+    let results = store.search_by_vector("ns", &[1.0, 0.0, 0.0], 5).unwrap();
+
+    assert_eq!(results.len(), 1, "the row must still be returned");
+    assert_eq!(results[0].id, "bad");
+    assert!(
+        results[0].metadata.is_null(),
+        "invalid metadata json must fall back to Null"
+    );
+}
+
 #[test]
 fn search_by_vector_scores_correct() {
     let store = fake_store(3);