superlocalmemory 3.3.3 → 3.3.4

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "superlocalmemory",
-  "version": "3.3.3",
+  "version": "3.3.4",
   "description": "Information-geometric agent memory with mathematical guarantees. 4-channel retrieval, Fisher-Rao similarity, zero-LLM mode, EU AI Act compliant. Works with Claude, Cursor, Windsurf, and 17+ AI tools.",
   "keywords": [
     "ai-memory",

package/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "superlocalmemory"
-version = "3.3.3"
+version = "3.3.4"
 description = "Information-geometric agent memory with mathematical guarantees"
 readme = "README.md"
 license = {text = "MIT"}
@@ -98,6 +98,10 @@ testpaths = ["tests"]
 pythonpath = ["src"]
 markers = [
     "slow: marks tests as slow (deselect with '-m \"not slow\"')",
+    "ollama: marks tests that require a running Ollama instance",
+]
+filterwarnings = [
+    "ignore::DeprecationWarning:vaderSentiment",
 ]
 
 [tool.coverage.run]

package/src/superlocalmemory/cli/commands.py

@@ -113,6 +113,10 @@ def cmd_mode(args: Namespace) -> None:
         if (config.embedding.provider != updated.embedding.provider
                 or config.embedding.model_name != updated.embedding.model_name):
             print("  ⚠ Embedding model changed. Re-indexing will run on next recall.")
+
+        # V3.3.4: Warn if Mode C lacks cloud API key
+        if args.value == "c" and not updated.llm.api_key:
+            print("  ⚠ Mode C requires a cloud API key. Run: slm provider set")
     else:
         print(f"Current mode: {config.mode.value.upper()}")
 
@@ -356,12 +360,20 @@ def cmd_forget(args: Namespace) -> None:
             sys.exit(1)
         raise
 
+    dry_run = getattr(args, 'dry_run', False)
+
     if use_json:
         from superlocalmemory.cli.json_output import json_print
         if not matches:
             json_print("forget", data={"matched_count": 0, "deleted_count": 0, "matches": []})
             return
         match_items = [{"fact_id": f.fact_id, "content": f.content[:120]} for f in matches[:20]]
+        if dry_run:
+            json_print("forget", data={
+                "matched_count": len(matches), "deleted_count": 0,
+                "dry_run": True, "matches": match_items,
+            })
+            return
         if getattr(args, 'yes', False):
             for f in matches:
                 engine._db.delete_fact(f.fact_id)
@@ -387,6 +399,9 @@ def cmd_forget(args: Namespace) -> None:
     print(f"Found {len(matches)} matching memories:")
     for f in matches[:10]:
         print(f"  - {f.fact_id[:8]}... {f.content[:80]}")
+    if dry_run:
+        print(f"(dry run — {len(matches)} would be deleted)")
+        return
     if getattr(args, 'yes', False):
         for f in matches:
             engine._db.delete_fact(f.fact_id)
@@ -861,7 +876,8 @@ def cmd_trace(args: Namespace) -> None:
     try:
         config = SLMConfig.load()
         engine = MemoryEngine(config)
-        response = engine.recall(args.query, limit=5)
+        limit = getattr(args, 'limit', 10)
+        response = engine.recall(args.query, limit=limit)
     except Exception as exc:
         if use_json:
             from superlocalmemory.cli.json_output import json_print
@@ -1435,6 +1451,7 @@ def cmd_consolidate(args: Namespace) -> None:
 
     use_json = getattr(args, "json", False)
     cognitive = getattr(args, "cognitive", False)
+    dry_run = getattr(args, "dry_run", False)
     profile = getattr(args, "profile", "")
 
     if not cognitive:
@@ -1460,7 +1477,7 @@ def cmd_consolidate(args: Namespace) -> None:
         )
 
         consolidator = CognitiveConsolidator(db=engine._db)
-        result = consolidator.run_pipeline(pid)
+        result = consolidator.run_pipeline(pid, dry_run=dry_run)
     except Exception as exc:
         if use_json:
             from superlocalmemory.cli.json_output import json_print
@@ -1473,7 +1490,7 @@ def cmd_consolidate(args: Namespace) -> None:
     if use_json:
         from superlocalmemory.cli.json_output import json_print
         json_print("consolidate", data={
-            "clusters_found": result.clusters_found,
+            "clusters_processed": result.clusters_processed,
             "blocks_created": result.blocks_created,
             "facts_archived": result.facts_archived,
             "compression_ratio": round(result.compression_ratio, 3),
@@ -1484,7 +1501,7 @@ def cmd_consolidate(args: Namespace) -> None:
         return
 
     print("CCQ Cognitive Consolidation")
-    print(f"  Clusters found:     {result.clusters_found}")
+    print(f"  Clusters processed: {result.clusters_processed}")
     print(f"  Blocks created:     {result.blocks_created}")
     print(f"  Facts archived:     {result.facts_archived}")
     print(f"  Compression ratio:  {result.compression_ratio:.3f}")

package/src/superlocalmemory/cli/main.py

@@ -123,6 +123,7 @@ def main() -> None:
 
     forget_p = sub.add_parser("forget", help="Delete memories matching a query (fuzzy)")
     forget_p.add_argument("query", help="Query to match for deletion")
+    forget_p.add_argument("--dry-run", action="store_true", default=False, help="Preview matches without deleting")
     forget_p.add_argument("--yes", "-y", action="store_true", help="Skip confirmation prompt")
     forget_p.add_argument("--json", action="store_true", help="Output structured JSON (agent-native)")
 
@@ -151,6 +152,7 @@ def main() -> None:
 
     trace_p = sub.add_parser("trace", help="Recall with per-channel score breakdown")
     trace_p.add_argument("query", help="Search query")
+    trace_p.add_argument("--limit", type=int, default=10, help="Max results (default 10)")
     trace_p.add_argument("--json", action="store_true", help="Output structured JSON (agent-native)")
 
     # -- Diagnostics (continued) ----------------------------------------
@@ -217,6 +219,10 @@ def main() -> None:
         "--cognitive", action="store_true",
         help="Run CCQ cognitive consolidation",
     )
+    consolidate_p.add_argument(
+        "--dry-run", action="store_true", default=False,
+        help="Preview without applying",
+    )
     consolidate_p.add_argument("--profile", default="", help="Target profile")
     consolidate_p.add_argument("--json", action="store_true", help="Output structured JSON (agent-native)")
 

package/src/superlocalmemory/core/config.py

@@ -612,15 +612,15 @@ class SLMConfig:
 
         rt = data.get("retrieval", {})
         if rt:
-            # V3.3.2 migration: auto-enable ONNX cross-encoder.
-            # Pre-3.3.2 configs had use_cross_encoder=False because the
-            # PyTorch cross-encoder used ~1.5GB RAM. With ONNX backend
-            # (~200MB), it's now safe for all modes. Detect old configs
-            # by the absence of cross_encoder_backend field.
+            # V3.3.2 migration: add ONNX cross-encoder backend field.
+            # Pre-3.3.2 configs lacked cross_encoder_backend. Add it,
+            # but NEVER override an explicit use_cross_encoder setting.
+            # The user's explicit choice always wins.
             if "cross_encoder_backend" not in rt:
-                rt["use_cross_encoder"] = True
-                rt["cross_encoder_model"] = "cross-encoder/ms-marco-MiniLM-L-6-v2"
+                rt.setdefault("cross_encoder_model", "cross-encoder/ms-marco-MiniLM-L-6-v2")
                 rt["cross_encoder_backend"] = "onnx"
+                # Only auto-enable if user didn't explicitly set the field
+                rt.setdefault("use_cross_encoder", True)
             config.retrieval = RetrievalConfig(**{
                 k: v for k, v in rt.items()
                 if k in RetrievalConfig.__dataclass_fields__
@@ -768,6 +768,9 @@ class SLMConfig:
             )
 
         # Mode C — FULL POWER, UNRESTRICTED
+        # Don't carry over local-only providers (ollama) to cloud mode
+        c_provider = llm_provider if llm_provider not in ("ollama", "") else "openrouter"
+        c_model = llm_model if llm_provider not in ("ollama", "") else "anthropic/claude-sonnet-4"
         return cls(
             mode=mode,
             base_dir=_base,
@@ -779,8 +782,8 @@ class SLMConfig:
                 deployment_name=embedding_deployment,
             ),
             llm=LLMConfig(
-                provider=llm_provider or "azure",
-                model=llm_model or "gpt-4.1-mini",
+                provider=c_provider,
+                model=c_model,
                 api_key=llm_api_key,
                 api_base=llm_api_base,
             ),

package/src/superlocalmemory/core/worker_pool.py

@@ -142,8 +142,15 @@ class WorkerPool:
     # ------------------------------------------------------------------
 
     def _send(self, request: dict) -> dict:
-        """Send request to worker and get response. Thread-safe."""
-        return self._send_with_timeout(request, timeout=_REQUEST_TIMEOUT)
+        """Send request to worker and get response. Thread-safe.
+
+        Auto-retries once on worker death (idle timeout, crash).
+        """
+        resp = self._send_with_timeout(request, timeout=_REQUEST_TIMEOUT)
+        if not resp.get("ok") and "Worker" in resp.get("error", ""):
+            logger.info("Auto-restarting worker after failure, retrying request")
+            resp = self._send_with_timeout(request, timeout=_REQUEST_TIMEOUT)
+        return resp
 
     def _send_with_timeout(self, request: dict, timeout: float) -> dict:
         """Send request with configurable timeout. Thread-safe."""

package/src/superlocalmemory/encoding/cognitive_consolidator.py

@@ -214,11 +214,17 @@ class CognitiveConsolidator:
     # Public API
     # ------------------------------------------------------------------
 
-    def run_pipeline(self, profile_id: str) -> CCQPipelineResult:
+    def run_pipeline(
+        self, profile_id: str, dry_run: bool = False,
+    ) -> CCQPipelineResult:
         """Execute the full 6-step CCQ pipeline.
 
         Per-cluster error isolation: one cluster failure does NOT
         abort the pipeline (HR-07).
+
+        Args:
+            profile_id: Target profile.
+            dry_run: If True, identify clusters but don't apply changes.
         """
         # Step 1: Identify candidates
         candidates = self._step1_identify(profile_id)
@@ -230,6 +236,18 @@ class CognitiveConsolidator:
         if not clusters:
             return self._empty_result()
 
+        if dry_run:
+            return CCQPipelineResult(
+                clusters_processed=len(clusters),
+                blocks_created=0,
+                facts_archived=len(candidates),
+                total_bytes_before=0,
+                total_bytes_after=0,
+                compression_ratio=0.0,
+                audit_entries=(),
+                errors=(),
+            )
+
         # Process each cluster
         blocks_created = 0
         facts_archived = 0

package/src/superlocalmemory/encoding/emotional.py

@@ -30,8 +30,11 @@ def _get_vader():
     if _vader_analyzer is not None:
         return _vader_analyzer
     try:
-        from vaderSentiment.vaderSentiment import SentimentIntensityAnalyzer
-        _vader_analyzer = SentimentIntensityAnalyzer()
+        import warnings
+        with warnings.catch_warnings():
+            warnings.filterwarnings("ignore", category=DeprecationWarning, module="vaderSentiment")
+            from vaderSentiment.vaderSentiment import SentimentIntensityAnalyzer
+            _vader_analyzer = SentimentIntensityAnalyzer()
     except ImportError:
         logger.warning("vaderSentiment not installed — emotional tagging disabled")
         _vader_analyzer = None

package/src/superlocalmemory/encoding/entity_resolver.py

@@ -498,7 +498,7 @@ class EntityResolver:
                 max_tokens=256,
                 temperature=0.0,
             )
-            match = re.search(r"\{.*\}", response, re.DOTALL)
+            match = re.search(r"\{[^}]*\}", response)
             if not match:
                 return {}
 

package/src/superlocalmemory/math/polar_quant.py

@@ -103,7 +103,9 @@ class PolarQuantEncoder:
         """
         path_str = self._config.rotation_matrix_path
         if not path_str:
-            path_str = str(Path.home() / ".superlocalmemory" / "polar_rotation.npy")
+            path_str = str(
+                Path.home() / ".superlocalmemory" / f"polar_rotation_{self._d}.npy",
+            )
 
         path = Path(path_str)
 

package/src/superlocalmemory/retrieval/engine.py

@@ -83,6 +83,10 @@ class RetrievalEngine:
         self._bridge = bridge_discovery
         self._trust_scorer = trust_scorer
 
+        # V3.3.4: LRU cache for query embeddings (avoids redundant Ollama API calls)
+        self._query_embedding_cache: dict[str, list[float]] = {}
+        self._cache_max_size = 64
+
         # V3.2: ChannelRegistry for self-registration (Phase 0.5)
         from superlocalmemory.retrieval.channel_registry import ChannelRegistry
         self._registry = ChannelRegistry()
@@ -189,6 +193,21 @@ class RetrievalEngine:
 
     # -- Channel execution --------------------------------------------------
 
+    def _embed_query(self, query: str) -> list[float] | None:
+        """Embed query with LRU cache. Avoids redundant Ollama/API calls."""
+        if self._embedder is None:
+            return None
+        cached = self._query_embedding_cache.get(query)
+        if cached is not None:
+            return cached
+        emb = self._embedder.embed(query)
+        # Evict oldest if cache full
+        if len(self._query_embedding_cache) >= self._cache_max_size:
+            oldest = next(iter(self._query_embedding_cache))
+            del self._query_embedding_cache[oldest]
+        self._query_embedding_cache[query] = emb
+        return emb
+
     def _run_channels(
         self, query: str, profile_id: str, strat: QueryStrategy,
     ) -> dict[str, list[tuple[str, float]]]:
@@ -197,9 +216,20 @@ class RetrievalEngine:
         # Skip channels listed in disabled_channels (ablation support)
         disabled = set(self._config.disabled_channels)
 
-        if self._semantic is not None and self._embedder is not None and "semantic" not in disabled:
+        # V3.3.4: Embed query ONCE, reuse for semantic + hopfield channels
+        q_emb: list[float] | None = None
+        needs_embedding = (
+            (self._semantic is not None and "semantic" not in disabled)
+            or (self._hopfield is not None and "hopfield" not in disabled)
+        )
+        if needs_embedding:
+            try:
+                q_emb = self._embed_query(query)
+            except Exception as exc:
+                logger.warning("Query embedding failed: %s", exc)
+
+        if self._semantic is not None and q_emb is not None and "semantic" not in disabled:
             try:
-                q_emb = self._embedder.embed(query)
                 r = self._semantic.search(q_emb, profile_id, self._config.semantic_top_k)
                 if r:
                     out["semantic"] = r
@@ -231,13 +261,11 @@ class RetrievalEngine:
                 logger.warning("Temporal channel: %s", exc)
 
         # Phase G: Hopfield channel (6th) — energy-based pattern completion
-        if self._hopfield is not None and "hopfield" not in disabled:
+        if self._hopfield is not None and q_emb is not None and "hopfield" not in disabled:
             try:
-                q_emb = self._embedder.embed(query) if self._embedder else None
-                if q_emb is not None:
-                    r = self._hopfield.search(q_emb, profile_id, self._config.hopfield_top_k)
-                    if r:
-                        out["hopfield"] = r
+                r = self._hopfield.search(q_emb, profile_id, self._config.hopfield_top_k)
+                if r:
+                    out["hopfield"] = r
             except Exception as exc:
                 logger.warning("Hopfield channel: %s", exc)
 

package/src/superlocalmemory/retrieval/reranker.py

@@ -2,10 +2,13 @@
 # Licensed under the MIT License - see LICENSE file
 # Part of SuperLocalMemory V3 | https://qualixar.com | https://varunpratap.com
 
-"""SuperLocalMemory V3 — Cross-Encoder Reranker.
+"""SuperLocalMemory V3 — Cross-Encoder Reranker (Subprocess-Isolated).
 
-Scores (query, fact) pairs through a cross-encoder in a single forward
-pass. Lazy model loading, thread-safe via lock.
+V3.3.3: All PyTorch/ONNX model work runs in a SEPARATE subprocess.
+The main process (dashboard, MCP, CLI) NEVER imports torch and stays
+at ~60 MB. Same isolation pattern as EmbeddingService.
+
+The worker subprocess auto-kills after 2 minutes idle.
 
 Part of Qualixar | Author: Varun Pratap Bhardwaj
 License: MIT
@@ -13,49 +16,33 @@ License: MIT
 
 from __future__ import annotations
 
+import json
 import logging
-import platform
-import struct
+import os
+import subprocess
 import sys
 import threading
+import time
 from typing import Any
 
 from superlocalmemory.storage.models import AtomicFact
 
 logger = logging.getLogger(__name__)
 
-
-def _detect_onnx_variant() -> str:
-    """Auto-detect the best ONNX model variant for the current platform.
-
-    Returns the file_name parameter for CrossEncoder model_kwargs.
-    Platform detection:
-    - macOS ARM64 (Apple Silicon): qint8_arm64
-    - x86_64 with AVX2: quint8_avx2
-    - Everything else: default model.onnx (float32, works everywhere)
-    """
-    arch = platform.machine().lower()
-    is_64bit = struct.calcsize("P") * 8 == 64
-
-    if sys.platform == "darwin" and arch in ("arm64", "aarch64"):
-        return "onnx/model_qint8_arm64.onnx"
-
-    if arch in ("x86_64", "amd64") and is_64bit:
-        return "onnx/model_quint8_avx2.onnx"
-
-    return "onnx/model.onnx"
+_IDLE_TIMEOUT_SECONDS = 120  # 2 min → kill worker
+_SUBPROCESS_RESPONSE_TIMEOUT = 120  # 120s for ONNX cold start
+_WORKER_RECYCLE_AFTER = 500  # Recycle after N requests
 
 
 class CrossEncoderReranker:
     """Rerank candidate facts using a local cross-encoder model.
 
-    V3.3.2: Uses ONNX backend by default (~200MB) instead of full PyTorch
-    (~1.5GB). Three-tier fallback: ONNX → PyTorch → no reranking.
-    Auto-detects the optimal quantized ONNX variant per platform.
+    V3.3.3: SUBPROCESS-ISOLATED. The main process never imports
+    sentence_transformers or torch. All model work runs in a child
+    process via JSON over stdin/stdout.
 
-    When the model is unavailable (missing package, download failure,
-    offline environment), falls back to returning candidates in their
-    original score order — never crashes.
+    Non-blocking first-use: triggers background worker spawn, returns
+    fallback scores until worker is ready.
 
     Args:
         model_name: HuggingFace cross-encoder model identifier.
@@ -70,106 +57,207 @@ class CrossEncoderReranker:
     ) -> None:
         self._model_name = model_name
         self._backend = backend
-        self._model: Any = None
-        self._loaded = False
-        self._loading = False  # True while background load is in progress
-        self._active_backend: str = ""
+        self._worker_proc: subprocess.Popen | None = None
+        self._model_loaded = False  # True once worker confirms model is ready
+        self._worker_loading = False  # True while background warmup in progress
         self._lock = threading.Lock()
+        self._idle_timer: threading.Timer | None = None
+        self._request_count: int = 0
+
+        # Start background warmup immediately — worker loads model
+        # while the rest of init continues. First recall gets instant
+        # fallback; second recall uses the warm model.
+        self._start_background_warmup()
 
     # ------------------------------------------------------------------
-    # Lazy loading (non-blocking)
+    # Background warmup (non-blocking model load)
     # ------------------------------------------------------------------
 
-    def _ensure_model(self) -> None:
-        """Trigger model load in background (non-blocking).
-
-        On first call, starts loading in a background thread and returns
-        immediately. The model becomes available for subsequent calls
-        once loading completes. This prevents the 30s ONNX cold start
-        from blocking the first recall request.
+    def _start_background_warmup(self) -> None:
+        """Start worker and load model in background thread.
 
-        Three-tier fallback:
-        1. ONNX backend with platform-optimal quantization — ~100-200MB RAM
-        2. PyTorch backend (requires torch) — ~1.5GB RAM
-        3. No model (graceful degradation) — 0 RAM
+        Returns immediately. The worker loads the model in parallel
+        with the rest of engine initialization and the first recall.
         """
-        if self._loaded:
+        if self._worker_loading or self._model_loaded:
             return
+        self._worker_loading = True
+
+        def _warmup() -> None:
+            try:
+                self._ensure_worker()
+                if self._worker_proc is None:
+                    return
+                # Send load command and wait for response
+                req = json.dumps({
+                    "cmd": "load",
+                    "model_name": self._model_name,
+                    "backend": self._backend,
+                }) + "\n"
+                self._worker_proc.stdin.write(req)
+                self._worker_proc.stdin.flush()
+                resp_line = self._readline_with_timeout(
+                    self._worker_proc.stdout, _SUBPROCESS_RESPONSE_TIMEOUT,
+                )
+                if resp_line:
+                    resp = json.loads(resp_line)
+                    if resp.get("ok"):
+                        self._model_loaded = True
+                        logger.info(
+                            "Reranker worker warm (backend=%s)",
+                            resp.get("backend", "?"),
+                        )
+                        self._reset_idle_timer()
+            except Exception as exc:
+                logger.debug("Background reranker warmup failed: %s", exc)
+            finally:
+                self._worker_loading = False
+
+        t = threading.Thread(target=_warmup, daemon=True, name="ce-warmup")
+        t.start()
 
-        with self._lock:
-            if self._loaded or self._loading:
-                return
-            self._loading = True
+    # ------------------------------------------------------------------
+    # Worker management (mirrors EmbeddingService pattern)
+    # ------------------------------------------------------------------
 
-        # Load in background thread so first recall isn't blocked
-        loader = threading.Thread(
-            target=self._load_model, daemon=True, name="ce-loader",
-        )
-        loader.start()
+    def _ensure_worker(self) -> None:
+        """Spawn worker subprocess if not running. Non-blocking."""
+        if self._worker_proc is not None and self._worker_proc.poll() is None:
+            return
+        self._worker_proc = None
+        self._worker_ready = False
 
-    def _load_model(self) -> None:
-        """Actually load the model (runs in background thread)."""
+        worker_module = "superlocalmemory.core.reranker_worker"
         try:
-            from sentence_transformers import CrossEncoder
-
-            if self._backend == "onnx":
-                try:
-                    onnx_file = _detect_onnx_variant()
-                    model = CrossEncoder(
-                        self._model_name,
-                        backend="onnx",
-                        model_kwargs={"file_name": onnx_file},
-                    )
-                    self._model = model
-                    self._active_backend = "onnx"
-                    logger.info(
-                        "Cross-encoder loaded (ONNX %s): %s",
-                        onnx_file, self._model_name,
-                    )
-                except Exception as onnx_exc:
-                    logger.info(
-                        "ONNX backend unavailable (%s), falling back to PyTorch",
-                        onnx_exc,
-                    )
-                    model = CrossEncoder(self._model_name)
-                    self._model = model
-                    self._active_backend = "pytorch"
-                    logger.info(
-                        "Cross-encoder loaded (PyTorch fallback): %s",
-                        self._model_name,
-                    )
-            else:
-                model = CrossEncoder(self._model_name)
-                self._model = model
-                self._active_backend = "pytorch"
-                logger.info("Cross-encoder loaded: %s", self._model_name)
-        except ImportError:
-            logger.warning(
-                "sentence-transformers not installed; "
-                "cross-encoder reranking disabled"
+            env = {
+                **os.environ,
+                "CUDA_VISIBLE_DEVICES": "",
+                "PYTORCH_MPS_HIGH_WATERMARK_RATIO": "0.0",
+                "PYTORCH_MPS_MEM_LIMIT": "0",
+                "PYTORCH_ENABLE_MPS_FALLBACK": "1",
+                "TOKENIZERS_PARALLELISM": "false",
+                "TORCH_DEVICE": "cpu",
+            }
+            self._worker_proc = subprocess.Popen(
+                [sys.executable, "-m", worker_module],
+                stdin=subprocess.PIPE,
+                stdout=subprocess.PIPE,
+                stderr=subprocess.DEVNULL,
+                text=True,
+                bufsize=1,
+                env=env,
+                start_new_session=True,
             )
-        except OSError as exc:
-            logger.warning(
-                "Failed to load cross-encoder %s: %s",
-                self._model_name,
-                exc,
+            logger.info(
+                "Reranker worker spawned (PID %d)", self._worker_proc.pid,
             )
-        finally:
-            self._loaded = True
-            self._loading = False
+            self._worker_ready = True
+        except Exception as exc:
+            logger.warning("Failed to spawn reranker worker: %s", exc)
+            self._worker_proc = None
 
-    def _ensure_model_blocking(self) -> None:
-        """Load model synchronously (blocks until ready).
+    def _send_request(self, req: dict, timeout: float | None = None) -> dict | None:
+        """Send JSON request to worker, get response. Thread-safe.
 
-        Used by warmup and is_available where we need the model NOW.
+        Uses a short timeout (10s) for rerank requests since the model
+        should already be loaded by the background warmup. Uses the full
+        timeout only for explicit load/ping commands.
         """
-        if self._loaded:
-            return
+        effective_timeout = timeout or _SUBPROCESS_RESPONSE_TIMEOUT
+
         with self._lock:
-            if self._loaded:
-                return
-            self._loading = True
-        self._load_model()
+            if self._request_count >= _WORKER_RECYCLE_AFTER and self._worker_proc is not None:
+                logger.info("Recycling reranker worker after %d requests", self._request_count)
+                self._kill_worker()
+                self._model_loaded = False
+                self._request_count = 0
+
+            # Ensure worker is alive (re-spawn if crashed)
+            if self._worker_proc is None or self._worker_proc.poll() is not None:
+                self._ensure_worker()
+            if self._worker_proc is None:
+                return None
+
+            try:
+                msg = json.dumps(req) + "\n"
+                self._worker_proc.stdin.write(msg)
+                self._worker_proc.stdin.flush()
+
+                resp_line = self._readline_with_timeout(
+                    self._worker_proc.stdout,
+                    effective_timeout,
+                )
+                if not resp_line:
+                    logger.warning("Reranker worker timed out after %ds", effective_timeout)
+                    self._kill_worker()
+                    self._model_loaded = False
+                    return None
+
+                resp = json.loads(resp_line)
+                self._reset_idle_timer()
+                self._request_count += 1
+                return resp
+            except (BrokenPipeError, OSError, json.JSONDecodeError) as exc:
+                logger.warning("Reranker worker communication failed: %s", exc)
+                self._kill_worker()
+                self._model_loaded = False
+                return None
+
+    @staticmethod
+    def _readline_with_timeout(stream: Any, timeout_seconds: float) -> str:
+        """Read a line from stream with timeout. Returns '' on timeout."""
+        result_container: list[str] = []
+        error_container: list[Exception] = []
+
+        def _read() -> None:
+            try:
+                result_container.append(stream.readline())
+            except Exception as exc:
+                error_container.append(exc)
+
+        reader = threading.Thread(target=_read, daemon=True)
+        reader.start()
+        reader.join(timeout=timeout_seconds)
+
+        if reader.is_alive():
+            return ""
+        if error_container:
+            raise error_container[0]
+        return result_container[0] if result_container else ""
+
+    def _kill_worker(self) -> None:
+        """Terminate worker subprocess."""
+        if self._idle_timer is not None:
+            self._idle_timer.cancel()
+            self._idle_timer = None
+        if self._worker_proc is not None:
+            try:
+                self._worker_proc.stdin.write('{"cmd":"quit"}\n')
+                self._worker_proc.stdin.flush()
+                self._worker_proc.wait(timeout=3)
+            except Exception:
+                try:
+                    self._worker_proc.kill()
+                except Exception:
+                    pass
+            self._worker_proc = None
+            self._worker_ready = False
+
+    def _reset_idle_timer(self) -> None:
+        """Reset idle timer — kills worker after 2 min inactivity."""
+        if self._idle_timer is not None:
+            self._idle_timer.cancel()
+        self._idle_timer = threading.Timer(
+            _IDLE_TIMEOUT_SECONDS, self.unload,
+        )
+        self._idle_timer.daemon = True
+        self._idle_timer.start()
+
+    def unload(self) -> None:
+        """Kill the worker subprocess to free all memory."""
+        with self._lock:
+            self._kill_worker()
+            logger.info("CrossEncoderReranker: worker killed (idle timeout)")
 
     # ------------------------------------------------------------------
     # Public API
@@ -183,73 +271,62 @@ class CrossEncoderReranker:
     ) -> list[tuple[AtomicFact, float]]:
         """Rerank candidates by cross-encoder relevance.
 
-        Each (query, fact.content) pair is scored in a single forward
-        pass. Results are returned sorted by cross-encoder score.
-
-        When the model is unavailable, returns candidates sorted by
-        their existing score (graceful fallback).
-
-        Args:
-            query: User query text.
-            candidates: List of (AtomicFact, score) tuples from the
-                fusion stage.
-            top_k: Maximum results to return.
-
-        Returns:
-            Top-k (AtomicFact, cross_encoder_score) tuples, sorted
-            descending by cross-encoder score.
+        NON-BLOCKING: If the worker is still loading the model
+        (background warmup), returns candidates by existing score
+        immediately. Once the worker is warm, subsequent calls use
+        the cross-encoder. This means CLI first-call gets instant
+        results (without reranking), and MCP gets reranked results
+        (worker stays warm between calls).
         """
         if not candidates:
             return []
 
-        # Non-blocking: trigger background load if not yet started
-        self._ensure_model()
-
-        if self._model is None:
-            # Model not loaded yet (still loading in background or failed).
-            # Graceful fallback: return candidates sorted by existing score.
-            # Next recall will use the model once it's ready.
-            sorted_cands = sorted(
-                candidates, key=lambda x: x[1], reverse=True
-            )
+        # Non-blocking: if model isn't loaded yet, return fallback
+        if not self._model_loaded:
+            sorted_cands = sorted(candidates, key=lambda x: x[1], reverse=True)
             return sorted_cands[:top_k]
 
-        # Build (query, document) pairs for batch scoring
-        pairs: list[tuple[str, str]] = [
-            (query, fact.content) for fact, _ in candidates
-        ]
+        documents = [fact.content for fact, _ in candidates]
+
+        # Short timeout (10s) — model should already be loaded by warmup.
+        # If worker crashed or is still loading, fallback immediately.
+        resp = self._send_request({
+            "cmd": "rerank",
+            "query": query,
+            "documents": documents,
+        }, timeout=10.0)
 
-        scores = self._model.predict(pairs)
+        if resp is None or not resp.get("ok"):
+            # Fallback: return by existing score
+            sorted_cands = sorted(candidates, key=lambda x: x[1], reverse=True)
+            return sorted_cands[:top_k]
 
+        scores = resp["scores"]
         scored: list[tuple[AtomicFact, float]] = [
             (fact, float(score))
             for (fact, _), score in zip(candidates, scores)
         ]
-
         scored.sort(key=lambda x: x[1], reverse=True)
         return scored[:top_k]
 
     def score_pair(self, query: str, document: str) -> float:
-        """Score a single (query, document) pair.
-
-        Args:
-            query: Query text.
-            document: Document text.
-
-        Returns:
-            Relevance score (higher = more relevant). 0.0 if model
-            is unavailable.
-        """
-        self._ensure_model()
-
-        if self._model is None:
+        """Score a single (query, document) pair."""
+        resp = self._send_request({
+            "cmd": "score",
+            "query": query,
+            "document": document,
+            "model_name": self._model_name,
+            "backend": self._backend,
+        })
+
+        if resp is None or not resp.get("ok"):
             return 0.0
-
-        scores = self._model.predict([(query, document)])
-        return float(scores[0])
+        return float(resp.get("score", 0.0))
 
     @property
     def is_available(self) -> bool:
-        """Whether the cross-encoder model is loaded and ready."""
-        self._ensure_model_blocking()
-        return self._model is not None
+        """Whether the cross-encoder worker can be spawned."""
+        resp = self._send_request({"cmd": "ping"})
+        if resp is None:
+            return False
+        return resp.get("ok", False)

package/src/superlocalmemory/storage/embedding_migrator.py

@@ -36,11 +36,12 @@ _REINDEX_BATCH_SIZE = 50
 def _model_signature(config: SLMConfig) -> str:
     """Derive a deterministic signature from the active embedding config.
 
-    The signature combines provider + model_name + dimension so that
-    any change in embedding source is detected.
+    V3.3.4: Only model_name + dimension matter. Provider (sentence-transformers
+    vs ollama) doesn't change the embedding space when the model is the same.
+    This prevents spurious re-indexing when switching Mode A ↔ B.
     """
     emb = config.embedding
-    return f"{emb.provider}::{emb.model_name}::{emb.dimension}"
+    return f"{emb.model_name}::{emb.dimension}"
 
 
 def _read_stored_signature(config_dir: Path) -> str: