superlocalmemory 3.0.13 → 3.0.15

package/src/superlocalmemory/core/embeddings.py

@@ -8,16 +8,29 @@ Thread-safe, dimension-validated embedding with Fisher variance computation.
 Supports local (768-dim nomic) and cloud (3072-dim) models with EXPLICIT errors
 on dimension mismatch — NEVER silently falls back to a different dimension.
 
+Memory management: Forces CPU-only inference to prevent GPU memory accumulation.
+Auto-unloads model after idle timeout to keep long-running MCP servers lean.
+
 Part of Qualixar | Author: Varun Pratap Bhardwaj
 """
 
 from __future__ import annotations
 
 import logging
+import os
 import threading
 import time
 from typing import TYPE_CHECKING
 
+# Force CPU before any torch/sentence-transformers import.
+# On Apple Silicon, PyTorch defaults to Metal (MPS) which allocates 4-6 GB
+# of GPU shader buffers that grow over time and never get released.
+# On Windows/Linux with CUDA, similar GPU memory issues occur.
+# CPU-only keeps footprint under 1 GB (vs 6+ GB with GPU).
+os.environ.setdefault("CUDA_VISIBLE_DEVICES", "")
+os.environ.setdefault("PYTORCH_MPS_HIGH_WATERMARK_RATIO", "0.0")
+os.environ.setdefault("PYTORCH_ENABLE_MPS_FALLBACK", "1")
+
 import numpy as np
 
 if TYPE_CHECKING:
@@ -44,11 +57,16 @@ class DimensionMismatchError(RuntimeError):
     """
 
 
+_IDLE_TIMEOUT_SECONDS = 300  # 5 minutes — unload model after idle
+
+
 class EmbeddingService:
     """Thread-safe embedding service with strict dimension validation.
 
     Lazy-loads the underlying model on first embed call.
     Validates every output dimension against the configured expectation.
+    Auto-unloads after 5 minutes idle to keep MCP server memory low.
+    Forces CPU-only inference to prevent GPU memory accumulation.
     """
 
     def __init__(self, config: EmbeddingConfig) -> None:
@@ -57,6 +75,8 @@ class EmbeddingService:
         self._lock = threading.Lock()
         self._loaded = False
         self._available = True  # Set False if model can't load
+        self._last_used: float = 0.0
+        self._idle_timer: threading.Timer | None = None
 
     @property
     def is_available(self) -> bool:
@@ -65,6 +85,32 @@ class EmbeddingService:
             self._ensure_loaded()
         return self._available and self._model is not None
 
+    def unload(self) -> None:
+        """Explicitly unload the model to free memory.
+
+        Called automatically after idle timeout, or manually for cleanup.
+        The model will lazy-reload on next embed call.
+        """
+        with self._lock:
+            if self._model is not None:
+                del self._model
+                self._model = None
+                self._loaded = False
+                import gc
+                gc.collect()
+                logger.info("EmbeddingService: model unloaded (idle timeout)")
+
+    def _reset_idle_timer(self) -> None:
+        """Reset the idle unload timer after each use."""
+        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()
+        self._last_used = time.time()
+
     # ------------------------------------------------------------------
     # Public API
     # ------------------------------------------------------------------
@@ -91,6 +137,7 @@ class EmbeddingService:
             return None
         vec = self._encode_single(text)
         self._validate_dimension(vec)
+        self._reset_idle_timer()
         return vec.tolist()
 
     def embed_batch(self, texts: list[str]) -> list[list[float]]:
@@ -115,6 +162,7 @@ class EmbeddingService:
         vectors = self._encode_batch(texts)
         for vec in vectors:
             self._validate_dimension(vec)
+        self._reset_idle_timer()
         return [v.tolist() for v in vectors]
 
     def compute_fisher_params(
@@ -185,7 +233,13 @@ class EmbeddingService:
             self._loaded = True
 
     def _load_local_model(self) -> None:
-        """Load sentence-transformers model for local embedding."""
+        """Load sentence-transformers model for local embedding.
+
+        Forces CPU device to prevent GPU memory accumulation:
+        - Apple Silicon MPS: allocates 4-6 GB Metal shader buffers
+        - NVIDIA CUDA: allocates GPU VRAM that never releases
+        - CPU-only: stable ~880 MB footprint, no growth over time
+        """
         try:
             from sentence_transformers import SentenceTransformer
         except ImportError:
@@ -199,6 +253,7 @@ class EmbeddingService:
             return
         model = SentenceTransformer(
             self._config.model_name, trust_remote_code=True,
+            device="cpu",
         )
         actual_dim = model.get_sentence_embedding_dimension()
         if actual_dim != self._config.dimension:
@@ -208,7 +263,7 @@ class EmbeddingService:
             )
         self._model = model
         logger.info(
-            "EmbeddingService: local model loaded (%s, %d-dim)",
+            "EmbeddingService: local model loaded (%s, %d-dim, device=cpu)",
             self._config.model_name,
             actual_dim,
         )

package/src/superlocalmemory/storage/schema.py

@@ -216,6 +216,28 @@ CREATE INDEX IF NOT EXISTS idx_facts_interval
 # FTS5 virtual table on atomic_facts for full-text search
 # ---------------------------------------------------------------------------
 
+_SQL_V2_MIGRATION_CLEANUP: Final[str] = """
+-- Clean up stale V2 triggers that fire on active tables but reference
+-- renamed backup FTS tables. The V2→V3 migration renames tables via
+-- ALTER TABLE RENAME, which auto-updates trigger bodies to reference
+-- _v2_bak_* tables but leaves FTS5 delete-command column names stale.
+-- This causes: "table _v2_bak_*_fts has no column named *_fts"
+
+-- Drop V2-era triggers on memories table (memories_ai/ad/au)
+DROP TRIGGER IF EXISTS memories_ai;
+DROP TRIGGER IF EXISTS memories_ad;
+DROP TRIGGER IF EXISTS memories_au;
+
+-- Drop stale V3 triggers (may have been corrupted by V2 rename)
+DROP TRIGGER IF EXISTS atomic_facts_fts_insert;
+DROP TRIGGER IF EXISTS atomic_facts_fts_delete;
+DROP TRIGGER IF EXISTS atomic_facts_fts_update;
+
+-- Drop renamed V2 backup FTS virtual tables (and their shadow tables)
+DROP TABLE IF EXISTS "_v2_bak_atomic_facts_fts";
+DROP TABLE IF EXISTS "_v2_bak_memories_fts";
+"""
+
 _SQL_ATOMIC_FACTS_FTS: Final[str] = """
 CREATE VIRTUAL TABLE IF NOT EXISTS atomic_facts_fts
     USING fts5(
@@ -226,8 +248,11 @@ CREATE VIRTUAL TABLE IF NOT EXISTS atomic_facts_fts
     );
 
 -- Triggers to keep FTS in sync with atomic_facts.
+-- Always DROP+CREATE (not IF NOT EXISTS) to replace any stale triggers
+-- left by V2 migration.
+
 -- INSERT trigger
-CREATE TRIGGER IF NOT EXISTS atomic_facts_fts_insert
+CREATE TRIGGER atomic_facts_fts_insert
     AFTER INSERT ON atomic_facts
 BEGIN
     INSERT INTO atomic_facts_fts (rowid, fact_id, content)
@@ -235,7 +260,7 @@ BEGIN
 END;
 
 -- DELETE trigger
-CREATE TRIGGER IF NOT EXISTS atomic_facts_fts_delete
+CREATE TRIGGER atomic_facts_fts_delete
     AFTER DELETE ON atomic_facts
 BEGIN
     INSERT INTO atomic_facts_fts (atomic_facts_fts, rowid, fact_id, content)
@@ -243,7 +268,7 @@ BEGIN
 END;
 
 -- UPDATE trigger
-CREATE TRIGGER IF NOT EXISTS atomic_facts_fts_update
+CREATE TRIGGER atomic_facts_fts_update
     AFTER UPDATE OF content ON atomic_facts
 BEGIN
     INSERT INTO atomic_facts_fts (atomic_facts_fts, rowid, fact_id, content)
@@ -655,7 +680,9 @@ _DDL_ORDERED: Final[tuple[str, ...]] = (
     _SQL_COMPLIANCE_AUDIT,
     _SQL_BM25_TOKENS,
     _SQL_CONFIG,
-    # FTS5 must come after atomic_facts (content table)
+    # V2 migration cleanup — drop stale triggers/FTS before recreating
+    _SQL_V2_MIGRATION_CLEANUP,
+    # FTS5 must come after atomic_facts (content table) AND after cleanup
     _SQL_ATOMIC_FACTS_FTS,
 )
 

package/src/superlocalmemory/storage/v2_migrator.py

@@ -151,20 +151,33 @@ class V2Migrator:
         self._backup_db = self._v3_base / BACKUP_NAME
 
     def detect_v2(self) -> bool:
-        """Check if a V2 installation exists."""
+        """Check if a V2 installation exists.
+
+        Returns False if .claude-memory is a symlink (already migrated).
+        """
+        if self._v2_base.is_symlink():
+            return False
         return self._v2_db.exists() and self._v2_db.is_file()
 
     def is_already_migrated(self) -> bool:
-        """Check if migration has already been performed."""
+        """Check if migration has already been performed.
+
+        Detects migration by:
+        1. .claude-memory is a symlink to .superlocalmemory (definitive)
+        2. V3 schema tables exist in the V3 database
+        """
+        if self._v2_base.is_symlink():
+            return True
         if not self._v3_db.exists():
             return False
         try:
             conn = sqlite3.connect(str(self._v3_db))
             try:
-                tables = [r[0] for r in conn.execute(
+                tables = {r[0] for r in conn.execute(
                     "SELECT name FROM sqlite_master WHERE type='table'"
-                ).fetchall()]
-                return "semantic_facts" in tables and "v3_config" in tables
+                ).fetchall()}
+                # Check for actual V3 schema tables (not old migration markers)
+                return "atomic_facts" in tables and "canonical_entities" in tables
             finally:
                 conn.close()
         except Exception:
@@ -217,12 +230,12 @@ class V2Migrator:
 
         Returns dict with migration stats.
         """
-        if not self.detect_v2():
-            return {"success": False, "error": "No V2 installation found"}
-
         if self.is_already_migrated():
             return {"success": True, "message": "Already migrated"}
 
+        if not self.detect_v2():
+            return {"success": False, "error": "No V2 installation found"}
+
         stats = {"steps": []}
 
         try:
@@ -268,6 +281,19 @@ class V2Migrator:
             # Disable FK enforcement for migrated DBs (V2 schema is incompatible)
             conn.execute("PRAGMA foreign_keys=OFF")
 
+            # Drop ALL triggers before renaming tables.
+            # ALTER TABLE RENAME auto-updates trigger bodies but corrupts
+            # FTS5 delete-command column names, causing:
+            #   "table _v2_bak_*_fts has no column named *_fts"
+            v2_triggers = [r[0] for r in conn.execute(
+                "SELECT name FROM sqlite_master WHERE type='trigger'"
+            ).fetchall()]
+            for trigger in v2_triggers:
+                try:
+                    conn.execute(f'DROP TRIGGER IF EXISTS "{trigger}"')
+                except Exception:
+                    pass
+
             # Rename ALL tables with incompatible schemas (V2 + old alpha)
             # User data is in 'memories' table (already upgraded above)
             # Everything else is computed/derived and will be recreated by V3