superlocalmemory 3.4.32 → 3.4.34

package/CHANGELOG.md

@@ -10,6 +10,46 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 
 ---
 
+## [3.4.34] - 2026-04-25
+
+Fix: user's mode choice can no longer be silently overwritten.
+
+### Fixed
+- **Mode protection in `SLMConfig.save()`.** Any `save()` call that would
+  change the mode in `config.json` is now blocked unless the caller passes
+  `mode_change=True`. This prevents accidental mode resets when code creates
+  a fresh `SLMConfig()` (defaults to Mode A) and calls `save()` to persist
+  an unrelated field change. A warning is logged when a silent mode change
+  is blocked.
+- **MCP `set_mode` preserves user settings.** Previously `set_mode` created
+  a fresh `SLMConfig.for_mode()` that lost all user customizations (LLM
+  provider, API keys, embedding config, active profile). Now carries forward
+  all settings from the existing config, matching the dashboard behavior.
+- All intentional mode-change paths (`slm mode`, MCP `set_mode`, dashboard
+  PUT `/api/v3/mode`, setup wizard) pass `mode_change=True`.
+
+---
+
+## [3.4.33] - 2026-04-25
+
+Fix: daemon leaked SQLite connections to learning.db via bandit threadlocals.
+
+### Fixed
+- **Bandit threadlocal connection leak.** `reward_proxy.settle_stale_plays`
+  creates a `ContextualBandit` that opens a threadlocal connection via
+  `_conn_for`. When called from `asyncio.to_thread` (bandit_loops.py,
+  every 60 s), each thread-pool thread kept its connection open for the
+  process lifetime. Over 24 h this accumulated 12+ leaked file descriptors
+  and ~100 MB of wasted SQLite page-cache RAM. New
+  `bandit.close_threadlocal_conn()` function, called in the
+  `settle_stale_plays` finally block, ensures pool threads release their
+  connections immediately.
+- **Corrected embedding worker memory comment.** The `~200MB footprint`
+  note was written for `all-MiniLM-L6-v2`; the default model
+  `nomic-ai/nomic-embed-text-v1.5` uses ~1.1 GB via ONNX.
+
+---
+
 ## [3.4.32] - 2026-04-24
 
 Fix: concurrent remembers no longer block recalls on the shared embedder.

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "superlocalmemory",
-  "version": "3.4.32",
+  "version": "3.4.34",
   "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 = "3.4.32"
+version = "3.4.34"
 description = "Information-geometric agent memory with mathematical guarantees"
 readme = "README.md"
 license = {text = "AGPL-3.0-or-later"}

package/src/superlocalmemory/__init__.py

@@ -1,3 +1,3 @@
 """SuperLocalMemory — information-geometric agent memory."""
 
-__version__ = "3.4.32"
+__version__ = "3.4.34"

package/src/superlocalmemory/cli/commands.py

@@ -639,7 +639,7 @@ def cmd_mode(args: Namespace) -> None:
                 llm_api_key=config.llm.api_key,
                 llm_api_base=config.llm.api_base,
             )
-            updated.save()
+            updated.save(mode_change=True)
             json_print("mode", data={
                 "previous_mode": old_mode, "current_mode": args.value.upper(),
             }, next_actions=[
@@ -661,7 +661,7 @@ def cmd_mode(args: Namespace) -> None:
             llm_api_key=config.llm.api_key,
             llm_api_base=config.llm.api_base,
         )
-        updated.save()
+        updated.save(mode_change=True)
         print(f"Mode set to: {args.value.upper()}")
 
         # V3.3: Check if embedding model changed — inform about re-indexing

package/src/superlocalmemory/cli/setup_wizard.py

@@ -335,7 +335,7 @@ def run_wizard(auto: bool = False) -> None:
     if choice == "c" and interactive:
         configure_provider(config)
     else:
-        config.save()
+        config.save(mode_change=True)
 
     mode_names = {"a": "Local Guardian", "b": "Smart Local", "c": "Full Power"}
     print(f"\n  ✓ Mode {choice.upper()} ({mode_names[choice]}) configured")
@@ -421,7 +421,7 @@ def run_wizard(auto: bool = False) -> None:
         config.daemon_idle_timeout = 0
         print("\n  ✓ 24/7 Always-On mode")
 
-    config.save()
+    config.save(mode_change=True)
 
     # -- Step 6: Mesh Communication (v3.4.3) --
     print()
@@ -441,7 +441,7 @@ def run_wizard(auto: bool = False) -> None:
         print("  Auto-enabling Mesh (non-interactive)")
 
     config.mesh_enabled = mesh_choice in ("", "y", "yes")
-    config.save()
+    config.save(mode_change=True)
     print(f"\n  ✓ Mesh {'enabled' if config.mesh_enabled else 'disabled'}")
 
     # -- Step 7: Ingestion Adapters (v3.4.3) --
@@ -502,7 +502,7 @@ def run_wizard(auto: bool = False) -> None:
         print("  Auto-enabling entity compilation (non-interactive)")
 
     config.entity_compilation_enabled = ec_choice in ("", "y", "yes")
-    config.save()
+    config.save(mode_change=True)
     print(f"\n  ✓ Entity compilation {'enabled' if config.entity_compilation_enabled else 'disabled'}")
 
     # -- Step 9: Skill Evolution (v3.4.11) --
@@ -661,7 +661,7 @@ def check_first_use(command: str) -> None:
             from superlocalmemory.core.config import SLMConfig
             from superlocalmemory.storage.models import Mode
             config = SLMConfig.for_mode(Mode.A)
-            config.save()
+            config.save(mode_change=True)
             _mark_complete()
         except Exception:
             pass
@@ -749,6 +749,6 @@ def configure_provider(config: object) -> None:
         llm_api_key=api_key,
         llm_api_base=preset["base_url"],
     )
-    updated.save()
+    updated.save(mode_change=True)
     print(f"  Provider: {provider_name}")
     print(f"  Model: {preset['model']}")

package/src/superlocalmemory/core/config.py

@@ -697,8 +697,25 @@ class SLMConfig:
 
         return config
 
-    def save(self, config_path: Path | None = None) -> None:
-        """Save config to JSON file."""
+    def save(
+        self,
+        config_path: Path | None = None,
+        *,
+        mode_change: bool = False,
+    ) -> None:
+        """Save config to JSON file.
+
+        v3.4.34: mode protection. If the existing config.json has a mode
+        that differs from ``self.mode`` and the caller did NOT pass
+        ``mode_change=True``, the EXISTING mode is preserved.  This
+        prevents accidental mode resets when code creates a fresh
+        ``SLMConfig()`` (defaults to Mode A) and calls ``save()`` to
+        persist an unrelated field change.
+
+        Callers that intentionally switch mode (``slm mode b``, the MCP
+        ``set_mode`` tool, the dashboard PUT ``/api/v3/mode``) MUST pass
+        ``mode_change=True``.
+        """
         import json
         path = config_path or (self.base_dir / "config.json")
         path.parent.mkdir(parents=True, exist_ok=True)
@@ -710,8 +727,19 @@ class SLMConfig:
             except (json.JSONDecodeError, OSError):
                 pass
 
+        # v3.4.34: mode protection — preserve user's mode unless explicitly changing
+        effective_mode = self.mode.value
+        existing_mode = existing.get("mode")
+        if existing_mode and existing_mode != effective_mode and not mode_change:
+            logger.warning(
+                "SLMConfig.save(): mode change blocked (%s → %s). "
+                "Pass mode_change=True to override. Preserving '%s'.",
+                existing_mode, effective_mode, existing_mode,
+            )
+            effective_mode = existing_mode
+
         data = {
-            "mode": self.mode.value,
+            "mode": effective_mode,
             "active_profile": self.active_profile,
             "llm": {
                 "provider": self.llm.provider,

package/src/superlocalmemory/core/embedding_worker.py

@@ -63,7 +63,7 @@ def _load_embedding_model(name: str) -> tuple:
     """
     from sentence_transformers import SentenceTransformer
 
-    # Tier 1: ONNX (stable memory, ~200MB footprint)
+    # Tier 1: ONNX (stable memory; ~1.1 GB for nomic-embed-text-v1.5)
     try:
         m = SentenceTransformer(name, backend="onnx", trust_remote_code=True)
         return m, "onnx"

package/src/superlocalmemory/learning/bandit.py

@@ -176,6 +176,23 @@ def _conn_for(db_path: Path) -> sqlite3.Connection:
     return conn
 
 
+def close_threadlocal_conn() -> None:
+    """Close the threadlocal bandit connection on the calling thread.
+
+    v3.4.33: background callers (asyncio.to_thread pool threads) MUST call
+    this after finishing bandit work.  Without it, each pool thread keeps a
+    leaked connection to learning.db for the process lifetime — observed as
+    12+ open file descriptors and ~100 MB wasted page-cache RAM.
+    """
+    if _holder.conn is not None:
+        try:
+            _holder.conn.close()
+        except sqlite3.Error:  # pragma: no cover
+            pass
+        _holder.conn = None
+        _holder.path = None
+
+
 def _now_iso() -> str:
     return datetime.now(timezone.utc).isoformat(timespec="seconds")
 
@@ -520,6 +537,7 @@ def retention_sweep(
 __all__ = (
     "BanditChoice",
     "ContextualBandit",
+    "close_threadlocal_conn",
     "compute_stratum",
     "current_time_bucket",
     "retention_sweep",

package/src/superlocalmemory/learning/reward_proxy.py

@@ -317,8 +317,13 @@ def settle_stale_plays(
                 memory_conn.close()
             except sqlite3.Error:  # pragma: no cover
                 pass
-        # Don't close a caller-owned bandit instance.
-        _ = owns_bandit
+        # v3.4.33: close the threadlocal bandit connection so pool threads
+        # from asyncio.to_thread don't leak file descriptors to learning.db.
+        try:
+            from superlocalmemory.learning.bandit import close_threadlocal_conn
+            close_threadlocal_conn()
+        except Exception:  # pragma: no cover — defensive
+            pass
 
     return settled
 

package/src/superlocalmemory/mcp/tools_v3.py

@@ -70,8 +70,20 @@ def register_v3_tools(server, get_engine: Callable) -> None:
 
             mode_enum = Mode(mode_lower)
             old_config = SLMConfig.load()
-            config = SLMConfig.for_mode(mode_enum)
-            config.save()
+            config = SLMConfig.for_mode(
+                mode_enum,
+                llm_provider=old_config.llm.provider,
+                llm_model=old_config.llm.model,
+                llm_api_key=old_config.llm.api_key,
+                llm_api_base=old_config.llm.api_base,
+                embedding_provider=old_config.embedding.provider,
+                embedding_endpoint=old_config.embedding.api_endpoint,
+                embedding_key=old_config.embedding.api_key,
+                embedding_model_name=old_config.embedding.model_name,
+                embedding_dimension=old_config.embedding.dimension,
+            )
+            config.active_profile = old_config.active_profile
+            config.save(mode_change=True)
 
             # V3.3: Check if embedding model changed — flag for re-indexing
             needs_reindex = (

package/src/superlocalmemory/server/routes/v3_api.py

@@ -136,7 +136,7 @@ async def set_mode(request: Request):
             embedding_dimension=old_config.embedding.dimension,
         )
         new_config.active_profile = old_config.active_profile
-        new_config.save()
+        new_config.save(mode_change=True)
 
         # Audit the change before we lose context — proves who/when/what.
         # Captures the phantom-write case where `for_mode(C)` auto-defaults