feat: archive.py uses embedding backend for semantic search

- MnemosyneArchive.__init__ accepts optional EmbeddingBackend
- Auto-detects best backend via get_embedding_backend()
- semantic_search uses embedding cosine similarity when available
- Falls back to Jaccard token similarity gracefully

nexus/mnemosyne/archive.py

@@ -13,6 +13,7 @@ from typing import Optional
 
 from nexus.mnemosyne.entry import ArchiveEntry, _compute_content_hash
 from nexus.mnemosyne.linker import HolographicLinker
+from nexus.mnemosyne.embeddings import get_embedding_backend, EmbeddingBackend
 
 _EXPORT_VERSION = "1"
 
@@ -24,10 +25,21 @@ class MnemosyneArchive:
     MemPalace (ChromaDB) for vector-semantic search.
     """
 
-    def __init__(self, archive_path: Optional[Path] = None):
+    def __init__(
+        self,
+        archive_path: Optional[Path] = None,
+        embedding_backend: Optional[EmbeddingBackend] = None,
+        auto_embed: bool = True,
+    ):
         self.path = archive_path or Path.home() / ".hermes" / "mnemosyne" / "archive.json"
         self.path.parent.mkdir(parents=True, exist_ok=True)
-        self.linker = HolographicLinker()
+        self._embedding_backend = embedding_backend
+        if embedding_backend is None and auto_embed:
+            try:
+                self._embedding_backend = get_embedding_backend()
+            except Exception:
+                self._embedding_backend = None
+        self.linker = HolographicLinker(embedding_backend=self._embedding_backend)
         self._entries: dict[str, ArchiveEntry] = {}
         self._load()
 
@@ -143,33 +155,51 @@ class MnemosyneArchive:
         return [e for _, e in scored[:limit]]
 
     def semantic_search(self, query: str, limit: int = 10, threshold: float = 0.05) -> list[ArchiveEntry]:
-        """Semantic search using holographic linker similarity.
+        """Semantic search using embeddings or holographic linker similarity.
 
-        Scores each entry by Jaccard similarity between query tokens and entry
-        tokens, then boosts entries with more inbound links (more "holographic").
-        Falls back to keyword search if no entries meet the similarity threshold.
+        With an embedding backend: cosine similarity between query vector and
+        entry vectors, boosted by inbound link count.
+        Without: Jaccard similarity on tokens with link boost.
+        Falls back to keyword search if nothing meets the threshold.
 
         Args:
             query: Natural language query string.
             limit: Maximum number of results to return.
-            threshold: Minimum Jaccard similarity to be considered a semantic match.
+            threshold: Minimum similarity score to include in results.
 
         Returns:
             List of ArchiveEntry sorted by combined relevance score, descending.
         """
-        query_tokens = HolographicLinker._tokenize(query)
-        if not query_tokens:
-            return []
-
-        # Count inbound links for each entry (how many entries link TO this one)
+        # Count inbound links for link-boost
         inbound: dict[str, int] = {eid: 0 for eid in self._entries}
         for entry in self._entries.values():
             for linked_id in entry.links:
                 if linked_id in inbound:
                     inbound[linked_id] += 1
-
         max_inbound = max(inbound.values(), default=1) or 1
 
+        # Try embedding-based search first
+        if self._embedding_backend:
+            query_vec = self._embedding_backend.embed(query)
+            if query_vec:
+                scored = []
+                for entry in self._entries.values():
+                    text = f"{entry.title} {entry.content} {' '.join(entry.topics)}"
+                    entry_vec = self._embedding_backend.embed(text)
+                    if not entry_vec:
+                        continue
+                    sim = self._embedding_backend.similarity(query_vec, entry_vec)
+                    if sim >= threshold:
+                        link_boost = inbound[entry.id] / max_inbound * 0.15
+                        scored.append((sim + link_boost, entry))
+                if scored:
+                    scored.sort(key=lambda x: x[0], reverse=True)
+                    return [e for _, e in scored[:limit]]
+
+        # Fallback: Jaccard token similarity
+        query_tokens = HolographicLinker._tokenize(query)
+        if not query_tokens:
+            return []
         scored = []
         for entry in self._entries.values():
             entry_tokens = HolographicLinker._tokenize(f"{entry.title} {entry.content} {' '.join(entry.topics)}")
@@ -179,14 +209,13 @@ class MnemosyneArchive:
             union = query_tokens | entry_tokens
             jaccard = len(intersection) / len(union)
             if jaccard >= threshold:
-                link_boost = inbound[entry.id] / max_inbound * 0.2  # up to 20% boost
+                link_boost = inbound[entry.id] / max_inbound * 0.2
                 scored.append((jaccard + link_boost, entry))
-
         if scored:
             scored.sort(key=lambda x: x[0], reverse=True)
             return [e for _, e in scored[:limit]]
 
-        # Graceful fallback to keyword search
+        # Final fallback: keyword search
         return self.search(query, limit=limit)
 
     def get_linked(self, entry_id: str, depth: int = 1) -> list[ArchiveEntry]: