superlocalmemory 3.8.2__tar.gz → 3.8.3__tar.gz

{superlocalmemory-3.8.2/src/superlocalmemory.egg-info → superlocalmemory-3.8.3}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: superlocalmemory
-Version: 3.8.2
+Version: 3.8.3
 Summary: Local-first agent memory with auditable hybrid retrieval
 Author-email: Varun Pratap Bhardwaj <admin@superlocalmemory.com>
 License-Expression: AGPL-3.0-or-later
@@ -106,10 +106,10 @@ Dynamic: license-file
   </picture>
 </p>
 
-<h1 align="center">SuperLocalMemory V3.8.2</h1>
+<h1 align="center">SuperLocalMemory V3.8.3</h1>
 <p align="center"><strong>Enterprise-grade, local-first memory for AI agents and teams.</strong><br/>
 <em>A persistent, auditable long-term brain for your agents that runs on your own infrastructure — with multi-workspace isolation, role-based access, and GDPR + EU AI Act governance controls built in.</em></p>
-<p align="center"><code>v3.8.2</code> — one control plane: auditable retrieval · multi-scope memory (personal / shared / global) · Cache · Compress · trusted-peer Mesh · bounded loops — across CLI, MCP, dashboard, the <strong>Claude plugin</strong>, the <strong>Codex add-on</strong>, and documented IDE integrations.<br/>
+<p align="center"><code>v3.8.3</code> — one control plane: auditable retrieval · multi-scope memory (personal / shared / global) · Cache · Compress · trusted-peer Mesh · bounded loops — across CLI, MCP, dashboard, the <strong>Claude plugin</strong>, the <strong>Codex add-on</strong>, and documented IDE integrations.<br/>
 Proxy: <code>slm wrap claude</code> &nbsp;·&nbsp; MCP: add <code>slm_compress</code> to your config &nbsp;·&nbsp; Skill: zero-config</p>
 <p align="center"><strong>3 public research preprints</strong> (arXiv + Zenodo archives) · <a href="https://arxiv.org/abs/2603.02240">arXiv:2603.02240</a> · <a href="https://arxiv.org/abs/2603.14588">arXiv:2603.14588</a> · <a href="https://arxiv.org/abs/2604.04514">arXiv:2604.04514</a></p>
 

{superlocalmemory-3.8.2 → superlocalmemory-3.8.3}/README.md

@@ -5,10 +5,10 @@
   </picture>
 </p>
 
-<h1 align="center">SuperLocalMemory V3.8.2</h1>
+<h1 align="center">SuperLocalMemory V3.8.3</h1>
 <p align="center"><strong>Enterprise-grade, local-first memory for AI agents and teams.</strong><br/>
 <em>A persistent, auditable long-term brain for your agents that runs on your own infrastructure — with multi-workspace isolation, role-based access, and GDPR + EU AI Act governance controls built in.</em></p>
-<p align="center"><code>v3.8.2</code> — one control plane: auditable retrieval · multi-scope memory (personal / shared / global) · Cache · Compress · trusted-peer Mesh · bounded loops — across CLI, MCP, dashboard, the <strong>Claude plugin</strong>, the <strong>Codex add-on</strong>, and documented IDE integrations.<br/>
+<p align="center"><code>v3.8.3</code> — one control plane: auditable retrieval · multi-scope memory (personal / shared / global) · Cache · Compress · trusted-peer Mesh · bounded loops — across CLI, MCP, dashboard, the <strong>Claude plugin</strong>, the <strong>Codex add-on</strong>, and documented IDE integrations.<br/>
 Proxy: <code>slm wrap claude</code> &nbsp;·&nbsp; MCP: add <code>slm_compress</code> to your config &nbsp;·&nbsp; Skill: zero-config</p>
 <p align="center"><strong>3 public research preprints</strong> (arXiv + Zenodo archives) · <a href="https://arxiv.org/abs/2603.02240">arXiv:2603.02240</a> · <a href="https://arxiv.org/abs/2603.14588">arXiv:2603.14588</a> · <a href="https://arxiv.org/abs/2604.04514">arXiv:2604.04514</a></p>
 

{superlocalmemory-3.8.2 → superlocalmemory-3.8.3}/plugin-src/rules/AGENTS.md

@@ -128,4 +128,4 @@ When the SLM MCP server is unavailable, use these CLI equivalents:
 - **slm-optimize-advisor** — context compression and KV cache
 - **slm-governance-advisor** — scope/role compliance, retention policies, GDPR
 
-SuperLocalMemory v3.8.2 · Qualixar · AGPL-3.0-or-later
+SuperLocalMemory v3.8.3 · Qualixar · AGPL-3.0-or-later

{superlocalmemory-3.8.2 → superlocalmemory-3.8.3}/plugin-src/skills/slm-cache/SKILL.md

@@ -145,4 +145,4 @@ These subcommands control daemon-level cache settings. They do not read or write
 
 ---
 
-SuperLocalMemory v3.8.2 · Qualixar · AGPL-3.0-or-later
+SuperLocalMemory v3.8.3 · Qualixar · AGPL-3.0-or-later

{superlocalmemory-3.8.2 → superlocalmemory-3.8.3}/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 v3.8.2 · Qualixar · AGPL-3.0-or-later
+SuperLocalMemory v3.8.3 · Qualixar · AGPL-3.0-or-later

{superlocalmemory-3.8.2 → superlocalmemory-3.8.3}/plugin-src/skills/slm-graph/SKILL.md

@@ -311,4 +311,4 @@ profile. See `slm-profile` for the full profile switching workflow.
 
 ---
 
-SuperLocalMemory v3.8.2 · Qualixar · AGPL-3.0-or-later
+SuperLocalMemory v3.8.3 · Qualixar · AGPL-3.0-or-later

{superlocalmemory-3.8.2 → superlocalmemory-3.8.3}/plugin-src/skills/slm-recall/SKILL.md

@@ -236,4 +236,4 @@ before recalling, then switch back. See `slm-profile` for workspace switching.
 
 ---
 
-*SuperLocalMemory v3.8.2 · Qualixar · AGPL-3.0-or-later*
+*SuperLocalMemory v3.8.3 · Qualixar · AGPL-3.0-or-later*

{superlocalmemory-3.8.2 → superlocalmemory-3.8.3}/plugin-src/skills/slm-remember/SKILL.md

@@ -238,4 +238,4 @@ different workspace, use `switch_profile` first. See `slm-profile`.
 
 ---
 
-*SuperLocalMemory v3.8.2 · Qualixar · AGPL-3.0-or-later*
+*SuperLocalMemory v3.8.3 · Qualixar · AGPL-3.0-or-later*

{superlocalmemory-3.8.2 → superlocalmemory-3.8.3}/plugin-src/skills/slm-session/SKILL.md

@@ -227,4 +227,4 @@ explicitly and call `recall` with `include_global`/`include_shared` after
 
 ---
 
-*SuperLocalMemory v3.8.2 · Qualixar · AGPL-3.0-or-later*
+*SuperLocalMemory v3.8.3 · Qualixar · AGPL-3.0-or-later*

{superlocalmemory-3.8.2 → superlocalmemory-3.8.3}/plugin-src/skills/slm-status/SKILL.md

@@ -163,4 +163,4 @@ multi-profile setup. To switch the active profile, see `slm-profile`.
 
 ---
 
-SuperLocalMemory v3.8.2 · Qualixar · AGPL-3.0-or-later
+SuperLocalMemory v3.8.3 · Qualixar · AGPL-3.0-or-later

{superlocalmemory-3.8.2 → superlocalmemory-3.8.3}/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "superlocalmemory"
-version = "3.8.2"
+version = "3.8.3"
 description = "Local-first agent memory with auditable hybrid retrieval"
 readme = "README.md"
 license = "AGPL-3.0-or-later"

{superlocalmemory-3.8.2 → superlocalmemory-3.8.3}/src/superlocalmemory/__init__.py

@@ -32,7 +32,7 @@ if "OMP_NUM_THREADS" not in os.environ:
     os.environ["OMP_NUM_THREADS"] = "2"
 # ---------------------------------------------------------------------------
 
-__version__ = "3.8.2"
+__version__ = "3.8.3"
 
 _REQUIRED_VERSIONS = {
     "sentence_transformers": "5.3.0",

{superlocalmemory-3.8.2 → superlocalmemory-3.8.3}/src/superlocalmemory/server/routes/memories.py

@@ -19,6 +19,25 @@ from .helpers import (
 logger = logging.getLogger("superlocalmemory.routes.memories")
 router = APIRouter()
 
+# v3.8.3: GENEROUS latency budget for recall. SLM's value is quality recall
+# under heavy multi-agent load, so semantic recall is given ample time to
+# finish — the keyword fallback is a LAST-RESORT safety net for a genuine hang
+# (e.g. a wedged embedder), NOT an aggressive speed cutoff. Only if recall
+# exceeds this budget do we serve the fast keyword search so the caller ALWAYS
+# gets a result instead of hanging forever. Tune with SLM_SEARCH_RECALL_TIMEOUT_S.
+# The dashboard's fetch timeout is set ABOVE this so the browser waits for the
+# quality result rather than aborting early.
+_DEFAULT_RECALL_BUDGET_S = 25.0
+
+
+def _search_recall_timeout_s() -> float:
+    import os
+    try:
+        v = float(os.environ.get("SLM_SEARCH_RECALL_TIMEOUT_S", ""))
+        return v if v > 0 else _DEFAULT_RECALL_BUDGET_S
+    except (TypeError, ValueError):
+        return _DEFAULT_RECALL_BUDGET_S
+
 
 def _internal_error(detail: str = "Internal server error") -> HTTPException:
     """SEC-H-02: log the full traceback server-side; return a generic message.
@@ -526,35 +545,65 @@ async def search_memories(request: Request, body: SearchRequest):
             _window = getattr(body, "window", None) or ""
             if not _window and getattr(body, "date_from", None) and getattr(body, "date_to", None):
                 _window = f"{body.date_from}..{body.date_to}"
-            response = await loop.run_in_executor(
+            # v3.8.3: bound the synchronous recall. Under a concurrent
+            # maintenance pass or a busy embedder it can run tens of seconds
+            # and the browser aborts the fetch. If it exceeds the budget we
+            # fall through to the fast keyword search below, so the dashboard
+            # ALWAYS returns instead of failing with an abort.
+            _recall_future = loop.run_in_executor(
                 None,
                 lambda: engine.recall(
                     body.query, limit=body.limit, fast=True,
                     window=_window or None,
                 ),
             )
-            elapsed_ms = round((_time.monotonic() - t0) * 1000, 1)
-            from superlocalmemory.server.recall_serializer import (
-                recall_response_metadata,
-                serialize_recall_response,
-            )
-            results, no_confident_match = serialize_recall_response(
-                response,
-                limit=body.limit,
-                per_fact_max=300,
-                total_max=max(300, body.limit * 300),
-            )
-            return {
-                "query": body.query,
-                "results": results,
-                "total": len(results),
-                "query_type": getattr(response, "query_type", "semantic"),
-                "retrieval_time_ms": elapsed_ms,
-                "no_confident_match": no_confident_match,
-                **recall_response_metadata(response),
-            }
-
-        # Fallback: direct DB text search (engine not yet initialised)
+            # A run_in_executor thread cannot be cancelled, and wait_for() on it
+            # blocks until the thread finishes (defeating the timeout). So poll
+            # the future without blocking the event loop and give up at the
+            # deadline — the orphaned recall completes in the background and its
+            # result is discarded. This is what bounds dashboard-search latency.
+            _budget = _search_recall_timeout_s()
+            _deadline = loop.time() + _budget
+            while not _recall_future.done() and loop.time() < _deadline:
+                await asyncio.sleep(0.05)
+            if _recall_future.done():
+                response = _recall_future.result()
+            else:
+                # Ensure the orphaned future's eventual result/exception is
+                # retrieved so asyncio doesn't log "never retrieved".
+                _recall_future.add_done_callback(
+                    lambda f: (f.cancelled() or f.exception())
+                )
+                logger.warning(
+                    "search_memories: semantic recall exceeded %.0fs budget for "
+                    "%r — serving keyword fallback",
+                    _budget, (body.query or "")[:80],
+                )
+                response = None
+            if response is not None:
+                elapsed_ms = round((_time.monotonic() - t0) * 1000, 1)
+                from superlocalmemory.server.recall_serializer import (
+                    recall_response_metadata,
+                    serialize_recall_response,
+                )
+                results, no_confident_match = serialize_recall_response(
+                    response,
+                    limit=body.limit,
+                    per_fact_max=300,
+                    total_max=max(300, body.limit * 300),
+                )
+                return {
+                    "query": body.query,
+                    "results": results,
+                    "total": len(results),
+                    "query_type": getattr(response, "query_type", "semantic"),
+                    "retrieval_time_ms": elapsed_ms,
+                    "no_confident_match": no_confident_match,
+                    **recall_response_metadata(response),
+                }
+            # recall timed out — fall through to the fast keyword search.
+
+        # Fallback: direct DB text search (engine not ready OR recall over budget)
         conn = get_db_connection()
         conn.row_factory = dict_factory
         cursor = conn.cursor()

{superlocalmemory-3.8.2 → superlocalmemory-3.8.3}/src/superlocalmemory/server/unified_daemon.py

@@ -453,6 +453,61 @@ def _emit_event(
 import asyncio as _asyncio
 _recall_semaphore = _asyncio.Semaphore(3)
 
+
+def _recall_budget_s() -> float:
+    """Generous latency budget for a recall before the keyword fallback (v3.8.3).
+
+    SLM's value is quality recall under heavy multi-agent load, so semantic
+    recall is given ample time; the keyword fallback is a LAST-RESORT safety
+    net for a genuine hang (e.g. a wedged embedder), not a speed cutoff. Tune
+    with SLM_SEARCH_RECALL_TIMEOUT_S (shared with the dashboard search route).
+    """
+    import os
+    try:
+        v = float(os.environ.get("SLM_SEARCH_RECALL_TIMEOUT_S", ""))
+        return v if v > 0 else 25.0
+    except (TypeError, ValueError):
+        return 25.0
+
+
+def _recall_keyword_fallback(engine, query: str, limit: int) -> dict:
+    """Fast profile-scoped keyword (LIKE) fallback for /recall.
+
+    Used only when semantic recall exceeds its budget, so CLI/MCP callers get
+    a bounded response instead of hanging. Mirrors the dashboard /api/search
+    fallback shape (retrieval_mode=degraded_lexical).
+    """
+    results = []
+    try:
+        rows = engine._db.execute(
+            "SELECT fact_id, content, confidence FROM atomic_facts "
+            "WHERE profile_id = ? AND content LIKE ? "
+            "ORDER BY confidence DESC LIMIT ?",
+            (engine.profile_id, f"%{query}%", limit),
+        )
+        for pos, r in enumerate(rows, start=1):
+            d = dict(r)
+            results.append({
+                "fact_id": d.get("fact_id"),
+                "content": (d.get("content") or "")[:2400],
+                "score": None, "relevance_score": None, "ranking_score": None,
+                "confidence": d.get("confidence"),
+                "rank_position": pos,
+            })
+    except Exception as exc:
+        logger.warning("recall keyword fallback failed (non-fatal): %s", exc)
+    return {
+        "ok": True,
+        "query": query,
+        "query_type": "text_search",
+        "retrieval_mode": "degraded_lexical",
+        "degraded_reason": "recall_budget_exceeded",
+        "result_count": len(results),
+        "results": results,
+        "count": len(results),
+        "no_confident_match": True,
+    }
+
 # v3.4.52: Embedding model warm state. Set to True by the async pre-warm
 # thread once Ollama has loaded the embedding model. /health reports this
 # so MCP clients can wait for warm state before issuing recall calls.
@@ -3013,15 +3068,35 @@ def _register_daemon_routes(application: FastAPI) -> None:
         if not fast:
             await _recall_semaphore.acquire()
         try:
-            response = await asyncio.to_thread(
-                engine.recall,
-                search_query, limit=limit, session_id=effective_sid,
-                agent_id=recall_actor,
-                fast=fast,
-                include_global=include_global,
-                include_shared=include_shared,
-                window=window or None,
+            # v3.8.3: bound the recall so CLI/MCP callers never hang on a
+            # wedged embedder. Poll the executor future (which cannot be
+            # cancelled) without blocking the loop, and give quality recall a
+            # GENEROUS budget; only if it is exceeded do we serve the fast
+            # keyword fallback. The orphaned recall finishes in the background.
+            loop = asyncio.get_running_loop()
+            _rf = loop.run_in_executor(
+                None,
+                lambda: engine.recall(
+                    search_query, limit=limit, session_id=effective_sid,
+                    agent_id=recall_actor,
+                    fast=fast,
+                    include_global=include_global,
+                    include_shared=include_shared,
+                    window=window or None,
+                ),
             )
+            _budget = _recall_budget_s()
+            _deadline = loop.time() + _budget
+            while not _rf.done() and loop.time() < _deadline:
+                await asyncio.sleep(0.05)
+            if not _rf.done():
+                _rf.add_done_callback(lambda f: (f.cancelled() or f.exception()))
+                logger.warning(
+                    "recall: semantic recall exceeded %.0fs budget for %r — "
+                    "serving keyword fallback", _budget, (search_query or "")[:80],
+                )
+                return _recall_keyword_fallback(engine, search_query, limit)
+            response = _rf.result()
             # v3.4.26: return the same field shape as recall_worker so
             # MCP processes proxying through the daemon get recall_trace-
             # compatible data without a second round trip.

{superlocalmemory-3.8.2 → superlocalmemory-3.8.3}/src/superlocalmemory/ui/js/core.js

@@ -13,7 +13,12 @@
 // surfaces as a normal rejection and the UI can show a clear error.
 // ============================================================================
 
-window.SLM_FETCH_TIMEOUT_MS = 15000;
+// v3.8.3: 30s (> the server's 25s recall budget) so the browser waits for a
+// quality recall under heavy multi-agent load instead of aborting it early
+// ("signal is aborted without reason"). The server self-bounds recall and
+// falls back to keyword within its budget, so this only ever waits longer for
+// a genuinely stuck request. Per-call init.timeoutMs still overrides.
+window.SLM_FETCH_TIMEOUT_MS = 30000;
 window.SLM_INSTALL_TOKEN_KEY = 'slm_install_token';
 
 // B2 (3.7.9): the install token is kept in a private closure, never in

{superlocalmemory-3.8.2 → superlocalmemory-3.8.3/src/superlocalmemory.egg-info}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: superlocalmemory
-Version: 3.8.2
+Version: 3.8.3
 Summary: Local-first agent memory with auditable hybrid retrieval
 Author-email: Varun Pratap Bhardwaj <admin@superlocalmemory.com>
 License-Expression: AGPL-3.0-or-later
@@ -106,10 +106,10 @@ Dynamic: license-file
   </picture>
 </p>
 
-<h1 align="center">SuperLocalMemory V3.8.2</h1>
+<h1 align="center">SuperLocalMemory V3.8.3</h1>
 <p align="center"><strong>Enterprise-grade, local-first memory for AI agents and teams.</strong><br/>
 <em>A persistent, auditable long-term brain for your agents that runs on your own infrastructure — with multi-workspace isolation, role-based access, and GDPR + EU AI Act governance controls built in.</em></p>
-<p align="center"><code>v3.8.2</code> — one control plane: auditable retrieval · multi-scope memory (personal / shared / global) · Cache · Compress · trusted-peer Mesh · bounded loops — across CLI, MCP, dashboard, the <strong>Claude plugin</strong>, the <strong>Codex add-on</strong>, and documented IDE integrations.<br/>
+<p align="center"><code>v3.8.3</code> — one control plane: auditable retrieval · multi-scope memory (personal / shared / global) · Cache · Compress · trusted-peer Mesh · bounded loops — across CLI, MCP, dashboard, the <strong>Claude plugin</strong>, the <strong>Codex add-on</strong>, and documented IDE integrations.<br/>
 Proxy: <code>slm wrap claude</code> &nbsp;·&nbsp; MCP: add <code>slm_compress</code> to your config &nbsp;·&nbsp; Skill: zero-config</p>
 <p align="center"><strong>3 public research preprints</strong> (arXiv + Zenodo archives) · <a href="https://arxiv.org/abs/2603.02240">arXiv:2603.02240</a> · <a href="https://arxiv.org/abs/2603.14588">arXiv:2603.14588</a> · <a href="https://arxiv.org/abs/2604.04514">arXiv:2604.04514</a></p>