superlocalmemory 4.2.8__tar.gz → 4.2.9__tar.gz

{superlocalmemory-4.2.8/src/superlocalmemory.egg-info → superlocalmemory-4.2.9}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: superlocalmemory
-Version: 4.2.8
+Version: 4.2.9
 Summary: The governed memory brain for AI agents: one local memory for Claude Code, Codex, Cursor, Copilot, ChatGPT and any MCP client
 Author-email: Varun Pratap Bhardwaj <admin@superlocalmemory.com>
 License-Expression: AGPL-3.0-or-later

{superlocalmemory-4.2.8 → superlocalmemory-4.2.9}/plugin-src/agents/slm-governance-advisor.md

@@ -81,4 +81,4 @@ slm-scope · slm-governance · slm-profile · slm-remember · slm-recall · slm-
 # What NOT to do
 Never session_init twice; never forget without dry-run preview; never run `slm gdpr erase --yes` unasked; never store secrets; never bypass role checks; never claim an erasure succeeded without verifying via recall.
 
-SuperLocalMemory v4.2.8 · Qualixar · AGPL-3.0-or-later
+SuperLocalMemory v4.2.9 · Qualixar · AGPL-3.0-or-later

{superlocalmemory-4.2.8 → superlocalmemory-4.2.9}/plugin-src/agents/slm-loop-runner.md

@@ -79,4 +79,4 @@ assessment. The gate is the authority.
 
 ---
 
-SuperLocalMemory v4.2.8 · Qualixar · AGPL-3.0-or-later
+SuperLocalMemory v4.2.9 · Qualixar · AGPL-3.0-or-later

{superlocalmemory-4.2.8 → superlocalmemory-4.2.9}/plugin-src/agents/slm-memory-advisor.md

@@ -51,4 +51,4 @@ slm-recall · slm-remember · slm-session · slm-scope · slm-profile · slm-gov
 # What NOT to do
 Never session_init twice; never forget dry_run=False without reporting preview; never switch the active profile unasked; never dump a whole file into remember; never invent a memory; never claim "saved" without success:true / clean CLI exit; never bypass scope or governance restrictions.
 
-SuperLocalMemory v4.2.8 · Qualixar · AGPL-3.0-or-later
+SuperLocalMemory v4.2.9 · Qualixar · AGPL-3.0-or-later

{superlocalmemory-4.2.8 → superlocalmemory-4.2.9}/plugin-src/agents/slm-optimize-advisor.md

@@ -41,4 +41,4 @@ slm-compress · slm-cache · slm-status · slm-profile
 # What NOT to do
 Never compress code-for-edit/JSON-to-parse/<500 chars; never store secrets/ccr_ids; never let optimize failure block/alter the task; never claim a specific savings %; never hand a ccr_id to another agent.
 
-SuperLocalMemory v4.2.8 · Qualixar · AGPL-3.0-or-later
+SuperLocalMemory v4.2.9 · Qualixar · AGPL-3.0-or-later

{superlocalmemory-4.2.8 → superlocalmemory-4.2.9}/plugin-src/rules/AGENTS.md

@@ -187,4 +187,4 @@ Using this memory from a web assistant or another computer: see the slm-web-acce
 - **slm-governance-advisor** — scope/role compliance, retention and lifecycle, GDPR
 - **slm-loop-runner** — bounded, gate-verified loops (the `/slm-loop` command delegates to it)
 
-SuperLocalMemory v4.2.8 · Qualixar · AGPL-3.0-or-later
+SuperLocalMemory v4.2.9 · Qualixar · AGPL-3.0-or-later

{superlocalmemory-4.2.8 → superlocalmemory-4.2.9}/plugin-src/rules/CLAUDE.md.fragment

@@ -1,4 +1,4 @@
-<!-- BEGIN SuperLocalMemory v4.2.8 -->
+<!-- BEGIN SuperLocalMemory v4.2.9 -->
 
 ## SuperLocalMemory (SLM) — Agent Rules
 
@@ -51,6 +51,6 @@ slm-recall · slm-remember · slm-session · slm-status · slm-cache · slm-comp
 ### Subagents
 slm-memory-advisor (memory decisions, session hygiene, scope/profile guidance) · slm-optimize-advisor (context compression + KV cache) · slm-governance-advisor (scope/roles/compliance/GDPR) · slm-loop-runner (bounded, gate-verified loops)
 
-<!-- END SuperLocalMemory v4.2.8 -->
+<!-- END SuperLocalMemory v4.2.9 -->
 
-SuperLocalMemory v4.2.8 · Qualixar · AGPL-3.0-or-later
+SuperLocalMemory v4.2.9 · Qualixar · AGPL-3.0-or-later

{superlocalmemory-4.2.8 → superlocalmemory-4.2.9}/plugin-src/skills/slm-bot-memory/SKILL.md

@@ -178,4 +178,4 @@ store every local bot on that computer can reach.
 
 ---
 
-*SuperLocalMemory v4.2.8 · Qualixar · AGPL-3.0-or-later*
+*SuperLocalMemory v4.2.9 · Qualixar · AGPL-3.0-or-later*

{superlocalmemory-4.2.8 → superlocalmemory-4.2.9}/plugin-src/skills/slm-cache/SKILL.md

@@ -144,4 +144,4 @@ These subcommands control daemon-level cache settings and bulk removal. They do
 
 ---
 
-SuperLocalMemory v4.2.8 · Qualixar · AGPL-3.0-or-later
+SuperLocalMemory v4.2.9 · Qualixar · AGPL-3.0-or-later

{superlocalmemory-4.2.8 → superlocalmemory-4.2.9}/plugin-src/skills/slm-compress/SKILL.md

@@ -147,4 +147,4 @@ Content over 1 MB (1 000 000 bytes UTF-8) is processed but `reversible` is force
 
 ---
 
-SuperLocalMemory v4.2.8 · Qualixar · AGPL-3.0-or-later
+SuperLocalMemory v4.2.9 · Qualixar · AGPL-3.0-or-later

{superlocalmemory-4.2.8 → superlocalmemory-4.2.9}/plugin-src/skills/slm-getting-started-bot/SKILL.md

@@ -194,4 +194,4 @@ quality on a shared box.
 
 ---
 
-*SuperLocalMemory v4.2.8 · Qualixar · AGPL-3.0-or-later*
+*SuperLocalMemory v4.2.9 · Qualixar · AGPL-3.0-or-later*

{superlocalmemory-4.2.8 → superlocalmemory-4.2.9}/plugin-src/skills/slm-governance/SKILL.md

@@ -276,4 +276,4 @@ Before running any destructive or state-changing operation (`forget`,
 
 ---
 
-*SuperLocalMemory v4.2.8 · Qualixar · AGPL-3.0-or-later*
+*SuperLocalMemory v4.2.9 · Qualixar · AGPL-3.0-or-later*

{superlocalmemory-4.2.8 → superlocalmemory-4.2.9}/plugin-src/skills/slm-graph/SKILL.md

@@ -318,4 +318,4 @@ All are host-only: a remote caller cannot use them.
 
 ---
 
-SuperLocalMemory v4.2.8 · Qualixar · AGPL-3.0-or-later
+SuperLocalMemory v4.2.9 · Qualixar · AGPL-3.0-or-later

{superlocalmemory-4.2.8 → superlocalmemory-4.2.9}/plugin-src/skills/slm-loop/SKILL.md

@@ -171,4 +171,4 @@ Guides: `docs/skill-evolution.md`, `docs/bounded-loops-bridge.md`.
 
 ---
 
-SuperLocalMemory v4.2.8 · Qualixar · AGPL-3.0-or-later
+SuperLocalMemory v4.2.9 · Qualixar · AGPL-3.0-or-later

{superlocalmemory-4.2.8 → superlocalmemory-4.2.9}/plugin-src/skills/slm-media/SKILL.md

@@ -337,4 +337,4 @@ minutes (previous vectors are kept) and let them decide.
 
 ---
 
-*SuperLocalMemory v4.2.8 · Qualixar · AGPL-3.0-or-later*
+*SuperLocalMemory v4.2.9 · Qualixar · AGPL-3.0-or-later*

{superlocalmemory-4.2.8 → superlocalmemory-4.2.9}/plugin-src/skills/slm-mesh/SKILL.md

@@ -386,4 +386,4 @@ mesh availability.
 
 ---
 
-*SuperLocalMemory v4.2.8 · Qualixar · AGPL-3.0-or-later*
+*SuperLocalMemory v4.2.9 · Qualixar · AGPL-3.0-or-later*

{superlocalmemory-4.2.8 → superlocalmemory-4.2.9}/plugin-src/skills/slm-profile/SKILL.md

@@ -164,4 +164,4 @@ anything but `core` or `mesh`.
 
 ---
 
-*SuperLocalMemory v4.2.8 · Qualixar · AGPL-3.0-or-later*
+*SuperLocalMemory v4.2.9 · Qualixar · AGPL-3.0-or-later*

{superlocalmemory-4.2.8 → superlocalmemory-4.2.9}/plugin-src/skills/slm-recall/SKILL.md

@@ -515,4 +515,4 @@ See `slm-profile`.
 
 ---
 
-*SuperLocalMemory v4.2.8 · Qualixar · AGPL-3.0-or-later*
+*SuperLocalMemory v4.2.9 · Qualixar · AGPL-3.0-or-later*

{superlocalmemory-4.2.8 → superlocalmemory-4.2.9}/plugin-src/skills/slm-remember/SKILL.md

@@ -372,4 +372,4 @@ sessions on this machine, so use it only when the user asks to move. See
 
 ---
 
-*SuperLocalMemory v4.2.8 · Qualixar · AGPL-3.0-or-later*
+*SuperLocalMemory v4.2.9 · Qualixar · AGPL-3.0-or-later*

{superlocalmemory-4.2.8 → superlocalmemory-4.2.9}/plugin-src/skills/slm-scope/SKILL.md

@@ -180,4 +180,4 @@ to review the impact. See `slm-remember` for the full deletion discipline.
 
 ---
 
-*SuperLocalMemory v4.2.8 · Qualixar · AGPL-3.0-or-later*
+*SuperLocalMemory v4.2.9 · Qualixar · AGPL-3.0-or-later*

{superlocalmemory-4.2.8 → superlocalmemory-4.2.9}/plugin-src/skills/slm-session/SKILL.md

@@ -288,4 +288,4 @@ answered. To also surface shared or global facts, call `recall` with
 
 ---
 
-*SuperLocalMemory v4.2.8 · Qualixar · AGPL-3.0-or-later*
+*SuperLocalMemory v4.2.9 · Qualixar · AGPL-3.0-or-later*

{superlocalmemory-4.2.8 → superlocalmemory-4.2.9}/plugin-src/skills/slm-status/SKILL.md

@@ -228,4 +228,4 @@ the memory-engine upgrade is pending.
 
 ---
 
-SuperLocalMemory v4.2.8 · Qualixar · AGPL-3.0-or-later
+SuperLocalMemory v4.2.9 · Qualixar · AGPL-3.0-or-later

{superlocalmemory-4.2.8 → superlocalmemory-4.2.9}/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "superlocalmemory"
-version = "4.2.8"
+version = "4.2.9"
 description = "The governed memory brain for AI agents: one local memory for Claude Code, Codex, Cursor, Copilot, ChatGPT and any MCP client"
 readme = "README.md"
 license = "AGPL-3.0-or-later"

{superlocalmemory-4.2.8 → superlocalmemory-4.2.9}/src/superlocalmemory/__init__.py

@@ -32,7 +32,7 @@ if "OMP_NUM_THREADS" not in os.environ:
     os.environ["OMP_NUM_THREADS"] = "2"
 # ---------------------------------------------------------------------------
 
-__version__ = "4.2.8"
+__version__ = "4.2.9"
 
 _REQUIRED_VERSIONS = {
     "sentence_transformers": "5.6.1",

{superlocalmemory-4.2.8 → superlocalmemory-4.2.9}/src/superlocalmemory/cli/setup_wizard.py

@@ -656,6 +656,21 @@ def _install_autostart_service() -> bool:
 # ---------------------------------------------------------------------------
 # Mode C provider config (preserved from original)
 # ---------------------------------------------------------------------------
+def _named_provider_mode(provider_name: str, target_mode: str | None):
+    """The mode ``slm provider set <provider>`` saves: ``--mode`` when given,
+    else B for a local model (Ollama) and C for a hosted one.
+
+    4.2.9: named providers were always saved as Mode C and ``--mode`` was read
+    only for ``custom``, so ``slm provider set ollama --mode b`` left Mode C.
+    """
+    from superlocalmemory.storage.models import Mode
+
+    choice = (target_mode or ("b" if provider_name == "ollama" else "c")).strip().lower()
+    if choice not in ("b", "c"):
+        raise ValueError(f"--mode must be b or c for an LLM provider, not {target_mode!r}")
+    return Mode.B if choice == "b" else Mode.C
+
+
 def configure_provider(
     config: object,
     provider_name: str | None = None,
@@ -737,7 +752,7 @@ def configure_provider(
     # preset model as the default. Non-interactive runs (``slm provider set``,
     # CI) keep the preset model — unchanged zero-prompt behavior. Logic lives
     # in cli.model_choice (this file is already past the 800-line guideline).
-    chosen_model = preset["model"]
+    chosen_model = (model or "").strip() or preset["model"]
     if interactive_selection and is_interactive():
         from superlocalmemory.cli.model_choice import pick_mode_c_model
         chosen_model = pick_mode_c_model(
@@ -749,8 +764,9 @@ def configure_provider(
     # configured a local or promoted Cozo/Lance deployment.  Keep every
     # unrelated setting intact and change only the mode/provider contract.
     from superlocalmemory.core.config import LLMConfig
+    mode = _named_provider_mode(provider_name, target_mode)
     updated = config if isinstance(config, SLMConfig) else SLMConfig.load()
-    updated.mode = Mode.C
+    updated.mode = mode
     updated.llm = LLMConfig(
         provider=provider_name,
         model=chosen_model,
@@ -758,6 +774,10 @@ def configure_provider(
         api_base=preset["base_url"],
     )
     updated.save(mode_change=True)
-    SLMConfig.write_current_mode(Mode.C, updated.base_dir)
+    # 4.2.9: the per-mode file too, or the next `slm mode` switch restores the
+    # model this command just replaced.
+    updated.save(SLMConfig._mode_config_path(updated.base_dir, mode))
+    SLMConfig.write_current_mode(mode.value, updated.base_dir)
+    print(f"  Mode: {mode.value.upper()}")
     print(f"  Provider: {provider_name}")
     print(f"  Model: {chosen_model}")

{superlocalmemory-4.2.8 → superlocalmemory-4.2.9}/src/superlocalmemory/core/config.py

@@ -2280,7 +2280,6 @@ class SLMConfig:
         _base.mkdir(parents=True, exist_ok=True)
 
         old_config = cls.load(_base / "config.json")
-        old_mode = old_config.mode.value.lower()
         new_mode_val = _M(new_mode.lower())
 
         # 4.1.18: the answer check's settings get their one home before any
@@ -2289,10 +2288,11 @@ class SLMConfig:
         _answer_check_state.ensure(
             _base, seed=lambda: _answer_check_state.snapshot(old_config.retrieval))
 
-        # 1. Save current config to its per-mode file (preserve customizations)
-        if old_mode != new_mode.lower():
-            old_path = cls._mode_config_path(_base, old_config.mode)
-            old_config.save(old_path)
+        # 1. Save current config to its per-mode file (preserve customizations).
+        #    4.2.9: also when switching to the mode already active. Skipping it
+        #    then loaded a stale mode_<x>.json and reverted every change saved
+        #    since the last switch (dashboard, wizard, `slm provider set`).
+        old_config.save(cls._mode_config_path(_base, old_config.mode))
 
         # 2. Determine new config:
         #    a) mode_{new}.json exists → user had prior config for this mode, use it

{superlocalmemory-4.2.8 → superlocalmemory-4.2.9}/src/superlocalmemory/core/model_catalog.py

@@ -60,7 +60,7 @@ def _e(**kw) -> ModelEntry:
 
 LOCAL_LLMS: tuple[ModelEntry, ...] = (
     _e(id="gemma4:e2b", role="llm", provider="ollama", label="Gemma 4 E2B",
-       size_gb=7.2, min_ram_gb=16, recommended=True,
+       size_gb=7.2, min_ram_gb=16, recommended=False,
        advice="Installed local alternative for extraction. No automatic switch; "
               "run your source-fidelity checks before choosing it."),
     _e(id="gemma3:4b", role="llm", provider="ollama", label="Gemma 3 4B",
@@ -139,12 +139,19 @@ def find(model_id: str) -> ModelEntry | None:
     return None
 
 
+#: Above this share of changed facts in SLM's extraction test a model ranks
+#: below an untested one: it has been seen inventing what was never said.
+_UNFAITHFUL_RATE = 0.10
+
+
 def _rank(entry: ModelEntry) -> tuple[float, str]:
     """Lower is better. (damaged + 1) / (tested + 2): a small sample with no
-    damage does not outrank a large one with almost none."""
+    damage does not outrank a large one with almost none. Untested models
+    follow every faithful tested one; models caught inventing facts go last."""
     if entry.damaged is None or not entry.tested:
-        return (1.0, entry.id)  # untested sorts after every tested model
-    return ((entry.damaged + 1) / (entry.tested + 2), entry.id)
+        return (1.0, entry.id)
+    rate = (entry.damaged + 1) / (entry.tested + 2)
+    return (rate if rate <= _UNFAITHFUL_RATE else 1.0 + rate, entry.id)
 
 
 def recommend_local_llms(ram_gb: float | None, installed: list[str]) -> list[Recommendation]:
@@ -163,11 +170,7 @@ def recommend_local_llms(ram_gb: float | None, installed: list[str]) -> list[Rec
                 or ram_gb >= entry.min_ram_gb)
 
     known = {e.id: e for e in LOCAL_LLMS}
-
-    def preference(entry: ModelEntry):
-        if entry.id == 'gemma4:e2b':
-            return (False, True, *_rank(entry))
-        return _rank(entry)
+    preference = _rank
     out: list[Recommendation] = []
     for entry in sorted((known[m] for m in have_llm if m in known), key=preference):
         if fits(entry):

{superlocalmemory-4.2.8 → superlocalmemory-4.2.9}/src/superlocalmemory/encoding/auto_linker.py

@@ -122,6 +122,11 @@ class AutoLinker:
         A-MEM memory evolution: old facts gain new context from new links.
         Updates fact_context table ONLY (immutability preserved -- Rule 17).
         Non-blocking, fire-and-forget. Errors logged, never raised (Rule 19).
+
+        4.2.9: the keyword path in every mode. Mode B/C used to regenerate each
+        linked fact through the LLM - up to 10 calls per stored memory, from a
+        prompt that held only the old fact, whose ContextResult was bound into
+        SQL that SQLite rejects, so nothing was ever saved (since v3.2.2).
         """
         try:
             new_fact = self._db.get_fact(new_fact_id)
@@ -129,17 +134,8 @@ class AutoLinker:
                 return
 
             for old_fact_id in linked_fact_ids:
-                old_fact = self._db.get_fact(old_fact_id)
-                if not old_fact:
-                    continue
-
-                if self._mode == "a":
+                if self._db.get_fact(old_fact_id):
                     self._append_keyword(old_fact_id, new_fact, profile_id)
-                else:
-                    if self._context_gen:
-                        self._regenerate_context(
-                            old_fact_id, old_fact, new_fact, profile_id,
-                        )
         except Exception as exc:
             logger.debug(
                 "evolve_linked_facts failed (non-fatal): %s", exc,
@@ -258,7 +254,7 @@ class AutoLinker:
     def _append_keyword(
         self, fact_id: str, new_fact: AtomicFact, profile_id: str,
     ) -> None:
-        """Mode A: append new fact's first entity as keyword."""
+        """Append the new fact's first entity (or first word) as a keyword."""
         keyword = ""
         if new_fact.canonical_entities:
             keyword = new_fact.canonical_entities[0]
@@ -287,22 +283,3 @@ class AutoLinker:
                     )
         except Exception as exc:
             logger.debug("_append_keyword failed: %s", exc)
-
-    def _regenerate_context(
-        self,
-        old_fact_id: str,
-        old_fact: AtomicFact,
-        new_fact: AtomicFact,
-        profile_id: str,
-    ) -> None:
-        """Mode B/C: regenerate contextual_description for old fact."""
-        try:
-            if self._context_gen:
-                new_desc = self._context_gen.generate(old_fact, self._mode)
-                self._db.execute(
-                    "UPDATE fact_context SET contextual_description = ? "
-                    "WHERE fact_id = ? AND profile_id = ?",
-                    (new_desc, old_fact_id, profile_id),
-                )
-        except Exception as exc:
-            logger.debug("_regenerate_context failed: %s", exc)

{superlocalmemory-4.2.8 → superlocalmemory-4.2.9}/src/superlocalmemory/encoding/context_generator.py

@@ -28,6 +28,26 @@ from superlocalmemory.storage.models import AtomicFact, FactType, SignalType
 
 logger = logging.getLogger(__name__)
 
+#: One sentence plus 3-5 keywords; the backbone default (4,096) let a model
+#: run on. Measured answers are ~75 tokens.
+CONTEXT_MAX_TOKENS = 160
+
+
+def _json_object(text: str) -> dict:
+    """The JSON object in a model answer, with or without a ```json fence.
+
+    4.2.9: gemma4:e2b fences its answer; a bare json.loads rejected it and
+    every context fell back to rules. Raises ValueError when there is none.
+    """
+    body = (text or "").strip()
+    start, end = body.find("{"), body.rfind("}")
+    if start < 0 or end < start:
+        raise ValueError("no JSON object in the model answer")
+    parsed = json.loads(body[start:end + 1])
+    if not isinstance(parsed, dict):
+        raise ValueError("the model answer is not a JSON object")
+    return parsed
+
 
 @dataclass(frozen=True)
 class ContextResult:
@@ -161,15 +181,17 @@ class ContextGenerator:
         response = self._llm.generate(
             prompt,
             system="You are a precise memory analyst. Respond ONLY with valid JSON.",
+            max_tokens=CONTEXT_MAX_TOKENS,
+            think=False,
         )
 
         try:
-            parsed = json.loads(response)
+            parsed = _json_object(response)
             return ContextResult(
                 description=parsed.get("description", "")[:500],
                 keywords=parsed.get("keywords", [])[:10],
                 generated_by="ollama" if mode == "b" else "cloud",
             )
-        except (json.JSONDecodeError, KeyError, TypeError):
+        except (ValueError, KeyError, TypeError):
             logger.debug("LLM context response unparseable, falling back to rules")
             return self._rules_based(fact)

{superlocalmemory-4.2.8 → superlocalmemory-4.2.9}/src/superlocalmemory/llm/backbone.py

@@ -91,6 +91,10 @@ class LLMResponse:
 LLM_YIELD_MAX_S = 20.0
 
 
+def _env_flag(name: str) -> bool:
+    return os.environ.get(name, "").strip().lower() in {"1", "true", "yes"}
+
+
 def _yield_to_recall() -> None:
     """Background generation (Mode B extraction, summaries) waits while a recall runs.
 
@@ -127,6 +131,9 @@ class LLMBackbone:
         self._timeout = config.timeout_seconds
         self._default_temperature = config.temperature
         self._default_max_tokens = config.max_tokens
+        # Set once an Ollama server rejects the think field (HTTP 400), so an
+        # old server costs one extra request per process, not one per call.
+        self._think_unsupported = False
 
         # Resolve API key: config > environment variable. A custom api_base
         # with an empty api_key is treated as intentionally keyless so local
@@ -260,6 +267,7 @@ class LLMBackbone:
                         and payload.pop("think", None) is not None
                     ):
                         think_downgraded = True
+                        self._think_unsupported = True
                         logger.info(
                             "Ollama rejected think:false (HTTP 400); "
                             "retrying once with model defaults.",
@@ -371,23 +379,29 @@ class LLMBackbone:
                 "num_ctx": 4096,
             },
         }
-        # Only an explicit False is ever sent (generate(think=False));
-        # True/None omit the key and keep model defaults.
-        if think is False:
-            payload["think"] = False
-        # #128: thinking-capable models (qwen3.x family) answer inside
-        # message.thinking, leaving content empty. The _extract_text
-        # fallback plus the extractor's array recovery cover that by
-        # default; operators who prefer plain content-only answers can
-        # opt out of thinking entirely with SLM_OLLAMA_DISABLE_THINK=1.
-        # Opt-in only and omitted by default: thinking support varies
-        # across model families, so the key is never sent unasked.
-        if os.environ.get(
-            "SLM_OLLAMA_DISABLE_THINK", "",
-        ).strip().lower() in {"1", "true", "yes"}:
+        if self._ollama_think_off(think):
             payload["think"] = False
         return self._base_url, headers, payload
 
+    def _ollama_think_off(self, think: bool | None) -> bool:
+        """Whether this request carries ``think: false``.
+
+        4.2.9: off unless asked for. Every SLM prompt wants a short structured
+        answer; a thinking model's default reasoning costs ~6x the tokens
+        (gemma4:e2b: ~410 vs ~75 per context call) and, under a token cap,
+        can leave content empty. ``True`` is never sent (non-thinking models
+        reject it): it, or ``SLM_OLLAMA_ALLOW_THINK=1``, keeps the model's
+        default. ``SLM_OLLAMA_DISABLE_THINK=1`` (#128) still forces it off.
+        A server that rejected the field once is not asked again.
+        """
+        if _env_flag("SLM_OLLAMA_DISABLE_THINK"):
+            return not self._think_unsupported
+        if self._think_unsupported or think is True:
+            return False
+        if think is False:
+            return True
+        return not _env_flag("SLM_OLLAMA_ALLOW_THINK")
+
     def _build_anthropic(
         self, prompt: str, system: str, max_tokens: int, temperature: float,
     ) -> tuple[str, dict[str, str], dict]:

{superlocalmemory-4.2.8 → superlocalmemory-4.2.9}/src/superlocalmemory/server/routes/v3_api.py

@@ -305,6 +305,11 @@ async def set_mode(request: Request):
             if _refusal is not None:
                 return JSONResponse(_refusal, status_code=400)
 
+        # 4.2.9: keep the mode being left in its own mode_<x>.json first, as
+        # `slm mode` does; otherwise a later CLI switch back loads a stale
+        # profile and reverts the model or embedder chosen in that mode.
+        old_config.save(SLMConfig._mode_config_path(old_config.base_dir, old_config.mode))
+
         # Apply new mode's structural presets (retrieval, math, channel_weights)
         # by building a fresh template, then graft them onto the loaded config so
         # all user-tuned blocks (forgetting, injection, consolidation, scope, …)
@@ -490,6 +495,10 @@ async def set_full_config(request: Request):
             return value.strip() or None
 
         new_mode = (_nonblank("mode") or config.mode.value).lower()
+        if new_mode != old_mode:
+            # 4.2.9: keep the mode being left in its mode_<x>.json, as `slm mode`
+            # does, so switching back never restores a stale profile.
+            config.save(SLMConfig._mode_config_path(config.base_dir, config.mode))
         provider_input = _nonblank("provider")
         model_input = _nonblank("model")
         api_key_input = _nonblank("api_key")

{superlocalmemory-4.2.8 → superlocalmemory-4.2.9/src/superlocalmemory.egg-info}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: superlocalmemory
-Version: 4.2.8
+Version: 4.2.9
 Summary: The governed memory brain for AI agents: one local memory for Claude Code, Codex, Cursor, Copilot, ChatGPT and any MCP client
 Author-email: Varun Pratap Bhardwaj <admin@superlocalmemory.com>
 License-Expression: AGPL-3.0-or-later

{superlocalmemory-4.2.8 → superlocalmemory-4.2.9}/tests/test_dashboard_preserve_embedding.py

@@ -631,3 +631,42 @@ class TestForModeContract:
 
         assert config.embedding.model_name == "nomic-ai/nomic-embed-text-v1.5"
         assert config.embedding.dimension == 768
+
+    def test_dashboard_switch_keeps_the_mode_it_leaves(self, isolated_config):
+        """4.2.9: PUT /api/v3/mode never saved the mode being left to its
+        mode_<x>.json, so a later `slm mode b` loaded a stale profile and
+        reverted the model chosen in B."""
+        live = {**_BASE_CONFIG, "mode": "b",
+                "llm": {**_BASE_CONFIG["llm"], "api_key": "sk-b", "model": "fresh-model"}}
+        _write_config(isolated_config, live)
+        stale = {**live, "llm": {**live["llm"], "model": "stale-model"}}
+        (isolated_config / "mode_b.json").write_text(json.dumps(stale), encoding="utf-8")
+
+        from superlocalmemory.core.config import SLMConfig
+        from superlocalmemory.server.routes.v3_api import set_mode
+
+        asyncio.run(set_mode(_make_request({"mode": "c"})))
+        assert _read_config(isolated_config)["mode"] == "c"
+
+        back = SLMConfig.switch_mode("b", base_dir=isolated_config)
+        assert back.llm.model == "fresh-model"
+
+    def test_settings_save_with_mode_change_keeps_the_mode_it_leaves(self, isolated_config):
+        """4.2.9: POST /api/v3/mode/set switching B -> C left mode_b.json stale."""
+        live = {**_BASE_CONFIG, "mode": "b",
+                "llm": {**_BASE_CONFIG["llm"], "model": "fresh-model"}}
+        _write_config(isolated_config, live)
+        stale = {**live, "llm": {**live["llm"], "model": "stale-model"}}
+        (isolated_config / "mode_b.json").write_text(json.dumps(stale), encoding="utf-8")
+
+        from superlocalmemory.core.config import SLMConfig
+        from superlocalmemory.server.routes.v3_api import set_full_config
+
+        asyncio.run(set_full_config(_make_request({
+            "mode": "c", "provider": "openrouter", "model": "openai/gpt-6-luna",
+            "api_key": "sk-new-key", "base_url": "https://openrouter.example.com",
+        })))
+        assert _read_config(isolated_config)["mode"] == "c"
+
+        back = SLMConfig.switch_mode("b", base_dir=isolated_config)
+        assert back.llm.model == "fresh-model"