get-claudia 1.55.15 → 1.55.16

package/CHANGELOG.md

@@ -2,6 +2,19 @@
 
 All notable changes to Claudia will be documented in this file.
 
+## 1.55.16 (2026-03-18)
+
+### Reliability Fixes
+
+Three fixes for issues surfaced from daemon logs. All backward-compatible, no schema changes.
+
+- **Overnight jobs now fire after sleep** -- APScheduler's `BackgroundScheduler` now has `misfire_grace_time=14400` (4 hours) and `coalesce=True`. Previously, the default 1-second grace time meant every scheduled job (decay, backup, consolidation, vault sync) was silently skipped when a Mac slept through the 2am-3:15am window. Now jobs fire immediately on wake if missed within the last 4 hours, with multiple missed runs collapsed into one execution.
+- **Reduced log noise from summary memories** -- The content length warning threshold was raised from 500 to 800 chars. Legitimate summary-type memories (550-850 chars) no longer trigger "Long content" warnings. Hard truncation at 1000 chars is unchanged.
+- **Fuzzy entity dedup on write** -- `_ensure_entity()` and `_find_or_create_entity()` now perform a fuzzy pre-check (SequenceMatcher > 0.90) before creating new entities. Name variants like "Kris Krisko" vs "Kris Krisco" (ratio ~0.92) match the existing entity instead of creating a duplicate. Only compares entities of the same type, skips deleted entities.
+- **Expanded STOP_WORDS** -- Added ~55 common English words that spaCy misidentifies as entities ("drawn", "overall", "recently", "several", etc.). Prevents ghost entities from cluttering the graph.
+- **Person entities require 2+ words** -- Regex-extracted person entities must have at least two words (e.g., "First Last"). Single-word extractions like "Metal" or "Drawn" are rejected. spaCy-identified entities are unaffected.
+- 637 tests pass, 0 regressions, 22 new tests across 4 new test files.
+
 ## 1.55.15 (2026-03-18)
 
 - **Fix mixed-timezone datetime crash** -- The memory daemon could crash with `can't subtract offset-naive and offset-aware datetimes` when recall or consolidation queries hit records with timezone suffixes (e.g., `+00:00` from email or transcript timestamps). Added a shared `parse_naive()` utility that strips timezone info on parse, applied across 14 locations in 5 files (recall.py, consolidate.py, server.py, vault_sync.py, canvas_generator.py). Replaces the older `[:19]` string truncation workaround. 615 tests pass.

package/memory-daemon/claudia_memory/daemon/scheduler.py

@@ -27,7 +27,9 @@ class MemoryScheduler:
     """Manages scheduled memory maintenance tasks"""
 
     def __init__(self):
-        self.scheduler = BackgroundScheduler()
+        self.scheduler = BackgroundScheduler(
+            job_defaults={"misfire_grace_time": 14400, "coalesce": True}
+        )
         self.config = get_config()
         self._started = False
 

package/memory-daemon/claudia_memory/extraction/entity_extractor.py

@@ -188,50 +188,33 @@ class EntityExtractor:
 
     # Common non-entity words to filter out
     STOP_WORDS = {
-        "monday",
-        "tuesday",
-        "wednesday",
-        "thursday",
-        "friday",
-        "saturday",
-        "sunday",
-        "january",
-        "february",
-        "march",
-        "april",
-        "may",
-        "june",
-        "july",
-        "august",
-        "september",
-        "october",
-        "november",
-        "december",
-        "today",
-        "tomorrow",
-        "yesterday",
-        "morning",
-        "afternoon",
-        "evening",
-        "night",
-        "the",
-        "this",
-        "that",
-        "these",
-        "those",
-        "here",
-        "there",
-        "where",
-        "when",
-        "what",
-        "which",
-        "who",
-        "how",
-        "just",
-        "only",
-        "also",
-        "even",
-        "still",
+        # Days and months
+        "monday", "tuesday", "wednesday", "thursday", "friday",
+        "saturday", "sunday",
+        "january", "february", "march", "april", "may", "june",
+        "july", "august", "september", "october", "november", "december",
+        # Temporal
+        "today", "tomorrow", "yesterday",
+        "morning", "afternoon", "evening", "night",
+        # Pronouns and determiners
+        "the", "this", "that", "these", "those",
+        "here", "there", "where", "when", "what", "which", "who", "how",
+        # Adverbs
+        "just", "only", "also", "even", "still",
+        "recently", "nearly", "almost", "already", "rather",
+        "somewhat", "perhaps", "quite", "likely", "enough",
+        # Quantifiers and adjectives
+        "several", "various", "another", "certain",
+        "much", "many", "some", "most", "both",
+        "each", "every", "other", "such", "same",
+        "new", "old", "big", "long", "last", "next",
+        "good", "well", "nice", "overall", "drawn",
+        # Common verbs (past tense / short forms spaCy misidentifies)
+        "done", "made", "said", "went", "got",
+        "set", "put", "run", "let", "get",
+        # Common nouns too generic to be entities
+        "work", "part", "plan", "team", "data",
+        "note", "time", "home", "call", "open",
     }
 
     def __init__(self):
@@ -296,12 +279,15 @@ class EntityExtractor:
         """Extract entities using regex patterns"""
         entities = []
 
-        # Extract persons
+        # Extract persons (require at least 2 words to avoid ghost entities)
         for pattern in self.PERSON_PATTERNS:
             for match in re.finditer(pattern, text):
                 name = match.group(1) if match.lastindex else match.group(0)
                 canonical = self.canonical_name(name)
-                if canonical and len(canonical) > 1 and canonical not in self.STOP_WORDS:
+                if (canonical
+                        and len(canonical) > 1
+                        and canonical not in self.STOP_WORDS
+                        and len(canonical.split()) >= 2):
                     entities.append(
                         ExtractedEntity(
                             name=name,

package/memory-daemon/claudia_memory/services/guards.py

@@ -42,7 +42,7 @@ def validate_memory(
     Validate a memory before storage.
 
     Checks:
-    - Content length (warn >500, truncate >1000)
+    - Content length (warn >800, truncate >1000)
     - Commitment deadline detection via regex
     - Importance clamped to [0, 1]
     """
@@ -52,7 +52,7 @@ def validate_memory(
     if len(content) > 1000:
         result.warnings.append(f"Content truncated from {len(content)} to 1000 characters")
         result.adjustments["content"] = content[:1000]
-    elif len(content) > 500:
+    elif len(content) > 800:
         result.warnings.append(f"Long content ({len(content)} chars) -- consider breaking into multiple memories")
 
     # Importance clamping

package/memory-daemon/claudia_memory/services/remember.py

@@ -1697,6 +1697,11 @@ class RememberService:
         if alias_match:
             return alias_match["entity_id"]
 
+        # Fuzzy pre-check: find near-matches of the same type
+        fuzzy_match = self._fuzzy_find_entity(extracted.canonical_name, extracted.type)
+        if fuzzy_match:
+            return fuzzy_match
+
         # Create new entity
         return self.remember_entity(
             name=extracted.name,
@@ -1725,9 +1730,44 @@ class RememberService:
         if alias_match:
             return alias_match["entity_id"]
 
+        # Fuzzy pre-check: find near-matches of the same type
+        fuzzy_match = self._fuzzy_find_entity(canonical, entity_type)
+        if fuzzy_match:
+            return fuzzy_match
+
         # Create new
         return self.remember_entity(name=name, entity_type=entity_type)
 
+    def _fuzzy_find_entity(self, canonical: str, entity_type: str) -> Optional[int]:
+        """Find a near-match entity of the same type using fuzzy string matching.
+
+        Queries entities of the given type and returns the ID of the best match
+        if similarity > 0.90 (SequenceMatcher ratio). Returns None if no match.
+        """
+        from difflib import SequenceMatcher
+
+        candidates = self.db.execute(
+            "SELECT id, canonical_name FROM entities WHERE type = ? AND deleted_at IS NULL",
+            (entity_type,),
+            fetch=True,
+        ) or []
+
+        best_id = None
+        best_ratio = 0.0
+        for row in candidates:
+            ratio = SequenceMatcher(None, canonical, row["canonical_name"]).ratio()
+            if ratio > 0.90 and ratio > best_ratio:
+                best_ratio = ratio
+                best_id = row["id"]
+
+        if best_id is not None:
+            logger.info(
+                f"Fuzzy entity match: '{canonical}' matched existing entity id={best_id} "
+                f"(type={entity_type}, similarity={best_ratio:.2f})"
+            )
+
+        return best_id
+
     def _get_or_create_episode(self, source: Optional[str] = None) -> int:
         """Get current episode or create a new one"""
         # For now, create a new episode each time

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "get-claudia",
-  "version": "1.55.15",
+  "version": "1.55.16",
   "description": "An AI assistant who learns how you work.",
   "keywords": [
     "claudia",