Add memory hashing, versioning, and forget support

src/agents/judge.py

@@ -87,6 +87,15 @@ def _format_similar_block(
     return "\n".join(lines)
 
 
+def _active_memory_results(results: List[SearchResult], limit: int) -> List[SearchResult]:
+    active = [
+        result for result in results
+        if result.metadata.get("is_current") is not False
+        and not result.metadata.get("forgotten_at")
+    ]
+    return active[:limit]
+
+
 # ---------------------------------------------------------------------------
 # Type alias for the Neo4j event search callable that the pipeline injects.
 #
@@ -264,7 +273,7 @@ async def _search_one(item_str: str) -> tuple[str, List[SearchResult]]:
                     query_text=item_str,
                     filters=filters if filters else None,
                 )
-                return item_str, results
+                return item_str, _active_memory_results(results, self.top_k)
             except Exception as exc:
                 self.logger.warning(
                     "Vector search failed for '%s': %s", item_str[:60], exc
@@ -306,9 +315,9 @@ async def _lookup_one(idx: int, item_str: str) -> tuple[str, List[SearchResult]]
                 if search_fn is not None:
                     # search_by_metadata is sync — run in thread pool
                     results = await asyncio.to_thread(
-                        search_fn, filters=filters, top_k=self.top_k,
+                        search_fn, filters=filters, top_k=max(self.top_k * 5, 10),
                     )
-                    return item_str, results if results else []
+                    return item_str, _active_memory_results(results or [], self.top_k)
                 else:
                     self.logger.debug(
                         "Vector store has no search_by_metadata — "
@@ -318,7 +327,7 @@ async def _lookup_one(idx: int, item_str: str) -> tuple[str, List[SearchResult]]
                         query_text=item_str,
                         filters={"user_id": user_id, "domain": "profile"} if user_id else None,
                     )
-                    return item_str, results
+                    return item_str, _active_memory_results(results, self.top_k)
             except Exception as exc:
                 self.logger.warning(
                     "Profile metadata search failed for '%s': %s",
@@ -344,7 +353,11 @@ async def _search_vector_store(
 
         search_fn = getattr(self.vector_store, "search_by_text", None)
         if search_fn is not None:
-            return await search_fn(query_text, top_k=self.top_k, filters=filters)
+            return await search_fn(
+                query_text,
+                top_k=max(self.top_k * 5, 10),
+                filters=filters,
+            )
 
         self.logger.debug(
             "Vector store has no search_by_text — skipping search for this item."

src/pipelines/retrieval.py

@@ -42,6 +42,10 @@
 logger = logging.getLogger("xmem.pipelines.retrieval")
 
 
+def _is_active_memory(metadata: Dict[str, Any]) -> bool:
+    return metadata.get("is_current") is not False and not metadata.get("forgotten_at")
+
+
 # ═══════════════════════════════════════════════════════════════════════════
 # Tool schemas — These are the "function signatures" exposed to the LLM
 # ═══════════════════════════════════════════════════════════════════════════
@@ -417,7 +421,7 @@ async def _search_summary(
 
         results = await self.vector_store.search_by_text(
             query_text=query,
-            top_k=top_k,
+            top_k=max(top_k * 2, 10),
             filters={
                 "user_id": user_id,
                 "domain": "summary",
@@ -426,12 +430,16 @@ async def _search_summary(
 
         records = []
         for r in results:
+            if not _is_active_memory(r.metadata):
+                continue
             records.append(SourceRecord(
                 domain="summary",
                 content=r.content,
                 score=r.score,
                 metadata={"id": r.id, **r.metadata},
             ))
+            if len(records) >= top_k:
+                break
 
         logger.info("  → Summary [%s]: %d results", query, len(records))
         return records
@@ -507,6 +515,8 @@ def _fetch_profile_catalog(self, user_id: str):
         seen = set()
 
         for r in results:
+            if not _is_active_memory(r.metadata):
+                continue
             main_content = r.metadata.get("main_content", "")
             if not main_content or main_content in seen:
                 continue