claude-self-reflect 3.2.3 → 3.2.4

package/.claude/agents/documentation-writer.md

@@ -48,7 +48,7 @@ You are a technical documentation specialist for the Claude Self Reflect project
  * @param query - Natural language search query
  * @param options - Search configuration options
  * @param options.limit - Maximum results to return (default: 10)
- * @param options.threshold - Minimum similarity score 0-1 (default: 0.7)
+ * @param options.threshold - Minimum similarity score 0-1 (removed in v3.2.4 - uses natural scoring)
  * @param options.project - Filter by specific project name
  * @returns Promise resolving to array of search results
  * 

package/.claude/agents/qdrant-specialist.md

@@ -12,7 +12,7 @@ You are a Qdrant vector database specialist for the claude-self-reflect project.
 - Collections use per-project isolation: `conv_<md5_hash>_local` or `conv_<md5_hash>_voyage` naming
 - Project paths: ~/.claude/projects/-Users-{username}-projects-{project-name}/*.jsonl
 - Project name is extracted from path and MD5 hashed for collection naming
-- Cross-collection search enabled with 0.7 similarity threshold
+- Cross-collection search uses Qdrant's natural scoring (no artificial thresholds since v3.2.4)
 - Streaming importer detects file growth and processes new lines incrementally
 - MCP server expects collections to match project name MD5 hash
 
@@ -195,7 +195,7 @@ docker stats qdrant
 
 ## Project-Specific Rules
 - Always use Voyage AI embeddings for consistency
-- Maintain 0.7 similarity threshold as baseline
+- Use Qdrant's natural scoring (no artificial thresholds since v3.2.4)
 - Preserve per-project collection isolation
 - Do not grep JSONL files unless explicitly asked
 - Always verify the MCP integration works end-to-end

package/.claude/agents/reflection-specialist.md

@@ -128,7 +128,7 @@ Fast search that returns only the count and top result. Perfect for quick checks
 // Quick overview of matches
 {
   query: "authentication patterns",
-  min_score: 0.5,  // Optional, defaults to 0.7
+  min_score: 0.5,  // Optional (v3.2.4+ ignores this - uses natural scoring)
   project: "all"    // Optional, defaults to current project
 }
 ```
@@ -165,7 +165,7 @@ Pagination support for getting additional results after an initial search.
   query: "original search query",  // Must match original query
   offset: 3,                      // Skip first 3 results
   limit: 3,                       // Get next 3 results
-  min_score: 0.7,                 // Optional
+  min_score: 0.7,                 // Optional (v3.2.4+ ignores this)
   project: "all"                  // Optional
 }
 ```

package/.claude/agents/search-optimizer.md

@@ -9,7 +9,7 @@ You are a search optimization specialist for the claude-self-reflect project. Yo
 ## Project Context
 - Current baseline: 66.1% search accuracy with Voyage AI
 - Gemini comparison showed 70-77% accuracy but 50% slower
-- Default similarity threshold: 0.7
+- Search scoring: Uses Qdrant's natural scoring (no artificial thresholds as of v3.2.4)
 - Cross-collection search adds ~100ms overhead
 - 24+ projects with 10,165+ conversation chunks
 
@@ -71,9 +71,11 @@ python scripts/analyze-search-quality.py
 ### Threshold Tuning
 ```bash
 # Test different thresholds
-for threshold in 0.5 0.6 0.7 0.8 0.9; do
-  echo "Testing threshold: $threshold"
-  SIMILARITY_THRESHOLD=$threshold npm test
+# Note: As of v3.2.4, artificial thresholds removed
+# Focus on embedding model comparison instead
+for model in voyage openai gemini; do
+  echo "Testing model: $model"
+  EMBEDDING_MODEL=$model npm test
 done
 
 # Find optimal threshold
@@ -237,7 +239,7 @@ def calculate_mrr(queries, results):
 interface ABTestConfig {
   control: {
     model: 'voyage',
-    threshold: 0.7,
+    scoring: 'natural',
     limit: 10
   },
   variant: {
@@ -285,7 +287,7 @@ async function abTestSearch(query: string, userId: string) {
 ### Recommended Settings
 ```env
 # Search Configuration
-SIMILARITY_THRESHOLD=0.7
+# SIMILARITY_THRESHOLD removed in v3.2.4 - uses natural scoring
 SEARCH_LIMIT=10
 CROSS_COLLECTION_LIMIT=5
 
@@ -300,7 +302,7 @@ SAMPLE_RATE=0.1
 ```
 
 ## Project-Specific Rules
-- Maintain 0.7 similarity threshold as baseline
+- Use Qdrant's natural scoring (no artificial thresholds since v3.2.4)
 - Always compare against Voyage AI baseline (66.1%)
 - Consider search latency alongside accuracy
 - Test with real conversation data

package/README.md

@@ -116,11 +116,9 @@ Works with [Claude Code Statusline](https://github.com/sirmalloc/ccstatusline) -
 <summary><b>MCP Tools Available to Claude</b></summary>
 
 **Search & Memory Tools:**
-- `reflect_on_past` - Search past conversations using semantic similarity with time decay
+- `reflect_on_past` - Search past conversations using semantic similarity with time decay (supports quick/summary modes)
 - `store_reflection` - Store important insights or learnings for future reference
-- `quick_search` - Fast search returning only count and top result
-- `search_summary` - Get aggregated insights without individual details
-- `get_more_results` - Paginate through additional search results
+- `get_next_results` - Paginate through additional search results
 - `search_by_file` - Find conversations that analyzed specific files
 - `search_by_concept` - Search for conversations about development concepts
 - `get_full_conversation` - Retrieve complete JSONL conversation files (v2.8.8)
@@ -288,11 +286,11 @@ npm uninstall -g claude-self-reflect
 ## What's New
 
 <details>
-<summary>v2.8.8 - Latest Release</summary>
+<summary>v3.2.4 - Latest Release</summary>
 
-- **Full Conversation Access**: New `get_full_conversation` tool provides complete JSONL files instead of 200-char excerpts
-- **95% Value Increase**: Agents can now access entire conversations with full implementation details
-- **Direct File Access**: Returns absolute paths for efficient reading with standard tools
+- **CRITICAL: Search Threshold Removal**: Eliminated artificial 0.7+ thresholds that blocked broad searches like "docker", "MCP", "python"
+- **Shared Normalization Module**: Created centralized project name normalization preventing search failures
+- **Memory Decay Fixes**: Corrected mathematical errors in exponential decay calculation
 
 </details>
 

package/mcp-server/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "claude-self-reflect-mcp"
-version = "2.8.9"
+version = "2.8.10"
 description = "MCP server for Claude self-reflection with memory decay"
 # readme = "README.md"
 requires-python = ">=3.10"

package/mcp-server/src/project_resolver.py

@@ -6,11 +6,21 @@ Handles mapping between user-friendly names and internal collection names.
 import hashlib
 import logging
 import re
+import sys
 from pathlib import Path
 from typing import List, Dict, Optional, Set
 from time import time
 from qdrant_client import QdrantClient
 
+# Import from shared module for consistent normalization
+sys.path.insert(0, str(Path(__file__).parent.parent.parent))
+try:
+    from shared.normalization import normalize_project_name
+except ImportError:
+    # Fall back to creating local version if shared module not found
+    logging.warning("Could not import shared normalization module")
+    normalize_project_name = None
+
 logger = logging.getLogger(__name__)
 
 # Project discovery markers - common parent directories that indicate project roots
@@ -244,59 +254,31 @@ class ProjectResolver:
     def _normalize_project_name(self, project_path: str) -> str:
         """
         Normalize project name for consistent hashing.
-        Extracts the actual project name from various path formats.
+        Uses the shared normalization module to ensure consistency
+        with import scripts.
         """
+        # Use the shared normalization function if available
+        if normalize_project_name:
+            return normalize_project_name(project_path)
+        
+        # Fallback implementation - EXACT copy of shared module
         if not project_path:
             return ""
         
-        # Remove trailing slashes
-        project_path = project_path.rstrip('/')
+        path = Path(project_path.rstrip('/'))
         
-        # Handle Claude logs format (starts with dash)
-        if project_path.startswith('-'):
-            # Split on dashes but don't convert to path separators
-            # This preserves project names that contain dashes
-            path_str = project_path[1:]  # Remove leading dash
-            path_parts = path_str.split('-')  # Split on dashes, not path separators
-            
-            # Look for common project parent directories
-            project_parents = {'projects', 'code', 'Code', 'repos', 'repositories', 
-                              'dev', 'Development', 'work', 'src', 'github'}
-            
-            # Find the project name after a known parent directory
-            for i, part in enumerate(path_parts):
-                if part.lower() in project_parents and i + 1 < len(path_parts):
-                    # Return everything after the parent directory
-                    remaining = path_parts[i + 1:]
-                    
-                    # Use segment-based approach for complex paths
-                    # Return the most likely project name from remaining segments
-                    if remaining:
-                        # If it's a single segment, return it
-                        if len(remaining) == 1:
-                            return remaining[0]
-                        # For multiple segments, look for project-like patterns
-                        for r in remaining:
-                            r_lower = r.lower()
-                            # Prioritize segments with project indicators
-                            if any(ind in r_lower for ind in ['app', 'service', 'project', 'api', 'client']):
-                                return r
-                    
-                    # Otherwise join remaining parts
-                    return '-'.join(remaining)
-            
-            # Fallback: use the last component
-            return path_parts[-1] if path_parts else project_path
+        # Extract the final directory name
+        final_component = path.name
         
-        # For regular paths or simple names
-        path_obj = Path(project_path)
+        # If it's Claude's dash-separated format, extract project name
+        if final_component.startswith('-') and 'projects' in final_component:
+            # Find the last occurrence of 'projects-' to handle edge cases
+            idx = final_component.rfind('projects-')
+            if idx != -1:
+                return final_component[idx + len('projects-'):]
         
-        # If it's already a simple name, return it
-        if '/' not in project_path and '\\' not in project_path:
-            return project_path
-            
-        # Otherwise extract from path
-        return path_obj.name
+        # For regular paths, just return the directory name
+        return final_component if final_component else path.parent.name
     
     def _project_matches(self, stored_project: str, target_project: str) -> bool:
         """

package/mcp-server/src/server.py

@@ -10,10 +10,22 @@ import numpy as np
 import hashlib
 import time
 import logging
+import math
 from xml.sax.saxutils import escape
 
 from fastmcp import FastMCP, Context
-from .utils import normalize_project_name
+
+# Import from shared module for consistent normalization
+import sys
+sys.path.insert(0, str(Path(__file__).parent.parent.parent))
+try:
+    from shared.normalization import normalize_project_name
+except ImportError:
+    # Fall back to local utils if shared module not found
+    from .utils import normalize_project_name
+    import logging
+    logging.warning("Using legacy utils.normalize_project_name - shared module not found")
+
 from .project_resolver import ProjectResolver
 from pydantic import BaseModel, Field
 from qdrant_client import AsyncQdrantClient, models
@@ -571,7 +583,7 @@ async def reflect_on_past(
     ctx: Context,
     query: str = Field(description="The search query to find semantically similar conversations"),
     limit: int = Field(default=5, description="Maximum number of results to return"),
-    min_score: float = Field(default=0.7, description="Minimum similarity score (0-1)"),
+    min_score: float = Field(default=0.3, description="Minimum similarity score (0-1)"),
     use_decay: Union[int, str] = Field(default=-1, description="Apply time-based decay: 1=enable, 0=disable, -1=use environment default (accepts int or str)"),
     project: Optional[str] = Field(default=None, description="Search specific project only. If not provided, searches current project based on working directory. Use 'all' to search across all projects."),
     include_raw: bool = Field(default=False, description="Include raw Qdrant payload data for debugging (increases response size)"),
@@ -669,8 +681,10 @@ async def reflect_on_past(
         # Filter collections by project if not searching all
         project_collections = []  # Define at this scope for later use
         if target_project != 'all':
-            # Use ProjectResolver to find collections for this project
-            resolver = ProjectResolver(qdrant_client)
+            # Use ProjectResolver with sync client (resolver expects sync operations)
+            from qdrant_client import QdrantClient as SyncQdrantClient
+            sync_client = SyncQdrantClient(url=QDRANT_URL)
+            resolver = ProjectResolver(sync_client)
             project_collections = resolver.find_collections_for_project(target_project)
             
             if not project_collections:
@@ -739,33 +753,32 @@ async def reflect_on_past(
                     await ctx.debug(f"Using NATIVE Qdrant decay (new API) for {collection_name}")
                     
                     # Build the query with native Qdrant decay formula using newer API
-                    query_obj = Query(
-                        nearest=query_embedding,
-                        formula=Formula(
+                    # Convert half-life to seconds (Qdrant uses seconds for datetime)
+                    half_life_seconds = DECAY_SCALE_DAYS * 24 * 60 * 60
+                    
+                    # Build query using proper Python models as per Qdrant docs
+                    from qdrant_client import models
+                    
+                    query_obj = models.FormulaQuery(
+                        formula=models.SumExpression(
                             sum=[
-                                # Original similarity score
-                                Expression(variable="score"),
-                                # Decay boost term
-                                Expression(
-                                    mult=MultExpression(
-                                        mult=[
-                                            # Decay weight
-                                            Expression(constant=DECAY_WEIGHT),
-                                            # Exponential decay function
-                                            Expression(
-                                                exp_decay=DecayParamsExpression(
-                                                    # Use timestamp field for decay
-                                                    x=Expression(datetime_key="timestamp"),
-                                                    # Decay from current time (server-side)
-                                                    target=Expression(datetime="now"),
-                                                    # Scale in milliseconds
-                                                    scale=DECAY_SCALE_DAYS * 24 * 60 * 60 * 1000,
-                                                    # Standard exponential decay midpoint
-                                                    midpoint=0.5
-                                                )
+                                "$score",  # Original similarity score
+                                models.MultExpression(
+                                    mult=[
+                                        DECAY_WEIGHT,  # Weight multiplier
+                                        models.ExpDecayExpression(
+                                            exp_decay=models.DecayParamsExpression(
+                                                x=models.DatetimeKeyExpression(
+                                                    datetime_key="timestamp"  # Payload field with datetime
+                                                ),
+                                                target=models.DatetimeExpression(
+                                                    datetime="now"  # Current time on server
+                                                ),
+                                                scale=half_life_seconds,  # Scale in seconds
+                                                midpoint=0.5  # Half-life semantics
                                             )
-                                        ]
-                                    )
+                                        )
+                                    ]
                                 )
                             ]
                         )
@@ -776,36 +789,32 @@ async def reflect_on_past(
                         collection_name=collection_name,
                         query=query_obj,
                         limit=limit,
-                        score_threshold=min_score,
                         with_payload=True
+                        # No score_threshold - let Qdrant's decay formula handle relevance
                     )
                 elif should_use_decay and USE_NATIVE_DECAY and not NATIVE_DECAY_AVAILABLE:
                     # Use native Qdrant decay with older API
                     await ctx.debug(f"Using NATIVE Qdrant decay (legacy API) for {collection_name}")
                     
+                    # Convert half-life to seconds (Qdrant uses seconds for datetime)
+                    half_life_seconds = DECAY_SCALE_DAYS * 24 * 60 * 60
+                    
                     # Build the query with native Qdrant decay formula using older API
+                    # Use the same models but with FormulaQuery
                     query_obj = FormulaQuery(
                         nearest=query_embedding,
                         formula=SumExpression(
                             sum=[
-                                # Original similarity score
-                                'score',  # Variable expression can be a string
-                                # Decay boost term
+                                "$score",  # Original similarity score
                                 {
-                                    'mult': [
-                                        # Decay weight (constant as float)
-                                        DECAY_WEIGHT,
-                                        # Exponential decay function
+                                    "mult": [
+                                        DECAY_WEIGHT,  # Weight multiplier
                                         {
-                                            'exp_decay': DecayParamsExpression(
-                                                # Use timestamp field for decay
-                                                x=DatetimeKeyExpression(datetime_key='timestamp'),
-                                                # Decay from current time (server-side)
-                                                target=DatetimeExpression(datetime='now'),
-                                                # Scale in milliseconds
-                                                scale=DECAY_SCALE_DAYS * 24 * 60 * 60 * 1000,
-                                                # Standard exponential decay midpoint
-                                                midpoint=0.5
+                                            "exp_decay": DecayParamsExpression(
+                                                x=DatetimeKeyExpression(datetime_key="timestamp"),
+                                                target=DatetimeExpression(datetime="now"),
+                                                scale=half_life_seconds,  # Scale in seconds
+                                                midpoint=0.5  # Half-life semantics
                                             )
                                         }
                                     ]
@@ -819,8 +828,8 @@ async def reflect_on_past(
                         collection_name=collection_name,
                         query=query_obj,
                         limit=limit,
-                        score_threshold=min_score,
                         with_payload=True
+                        # No score_threshold - let Qdrant's decay formula handle relevance
                     )
                     
                     # Process results from native decay search
@@ -916,11 +925,14 @@ async def reflect_on_past(
                                     timestamp = timestamp.replace(tzinfo=timezone.utc)
                                 age_ms = (now - timestamp).total_seconds() * 1000
                                 
-                                # Calculate decay factor
-                                decay_factor = np.exp(-age_ms / scale_ms)
+                                # Calculate decay factor using proper half-life formula
+                                # For half-life H: decay = exp(-ln(2) * age / H)
+                                ln2 = math.log(2)
+                                decay_factor = math.exp(-ln2 * age_ms / scale_ms)
                                 
-                                # Apply decay formula
-                                adjusted_score = point.score + (DECAY_WEIGHT * decay_factor)
+                                # Apply multiplicative decay formula to keep scores bounded [0, 1]
+                                # adjusted = score * ((1 - weight) + weight * decay)
+                                adjusted_score = point.score * ((1 - DECAY_WEIGHT) + DECAY_WEIGHT * decay_factor)
                                 
                                 # Debug: show the calculation
                                 age_days = age_ms / (24 * 60 * 60 * 1000)
@@ -1001,12 +1013,13 @@ async def reflect_on_past(
                         ))
                 else:
                     # Standard search without decay
+                    # Let Qdrant handle scoring natively
                     results = await qdrant_client.search(
                         collection_name=collection_name,
                         query_vector=query_embedding,
                         limit=limit * 2,  # Get more results to account for filtering
-                        score_threshold=min_score * 0.9,  # Slightly lower threshold to catch v1 chunks
                         with_payload=True
+                        # No score_threshold - let Qdrant decide what's relevant
                     )
                     
                     for point in results:
@@ -1691,7 +1704,7 @@ async def store_reflection(
 async def quick_search(
     ctx: Context,
     query: str = Field(description="The search query to find semantically similar conversations"),
-    min_score: float = Field(default=0.7, description="Minimum similarity score (0-1)"),
+    min_score: float = Field(default=0.3, description="Minimum similarity score (0-1)"),
     project: Optional[str] = Field(default=None, description="Search specific project only. If not provided, searches current project based on working directory. Use 'all' to search across all projects.")
 ) -> str:
     """Quick search that returns only the count and top result for fast overview."""
@@ -1737,7 +1750,7 @@ async def get_more_results(
     query: str = Field(description="The original search query"),
     offset: int = Field(default=3, description="Number of results to skip (for pagination)"),
     limit: int = Field(default=3, description="Number of additional results to return"),
-    min_score: float = Field(default=0.7, description="Minimum similarity score (0-1)"),
+    min_score: float = Field(default=0.3, description="Minimum similarity score (0-1)"),
     project: Optional[str] = Field(default=None, description="Search specific project only")
 ) -> str:
     """Get additional search results after an initial search (pagination support)."""
@@ -1772,8 +1785,9 @@ async def search_by_file(
     collections = await get_all_collections() if not project else []
     
     if project and project != 'all':
-        # Filter collections for specific project
-        project_hash = hashlib.md5(project.encode()).hexdigest()[:8]
+        # Filter collections for specific project - normalize first!
+        normalized_project = normalize_project_name(project)
+        project_hash = hashlib.md5(normalized_project.encode()).hexdigest()[:8]
         collection_prefix = f"conv_{project_hash}_"
         collections = [c for c in await get_all_collections() if c.startswith(collection_prefix)]
     elif project == 'all':
@@ -2137,7 +2151,7 @@ async def get_next_results(
     query: str = Field(description="The original search query"),
     offset: int = Field(default=3, description="Number of results to skip (for pagination)"),
     limit: int = Field(default=3, description="Number of additional results to return"),
-    min_score: float = Field(default=0.7, description="Minimum similarity score (0-1)"),
+    min_score: float = Field(default=0.3, description="Minimum similarity score (0-1)"),
     project: Optional[str] = Field(default=None, description="Search specific project only")
 ) -> str:
     """Get additional search results after an initial search (pagination support)."""
@@ -2152,9 +2166,10 @@ async def get_next_results(
             # Search all collections if project is "all" or not specified
             collections = await get_all_collections()
         else:
-            # Search specific project
+            # Search specific project - normalize first!
             all_collections = await get_all_collections()
-            project_hash = hashlib.md5(project.encode()).hexdigest()[:8]
+            normalized_project = normalize_project_name(project)
+            project_hash = hashlib.md5(normalized_project.encode()).hexdigest()[:8]
             collections = [
                 c for c in all_collections 
                 if c.startswith(f"conv_{project_hash}_")
@@ -2196,9 +2211,12 @@ async def get_next_results(
                     if use_decay_bool and 'timestamp' in payload:
                         try:
                             timestamp = datetime.fromisoformat(payload['timestamp'].replace('Z', '+00:00'))
-                            age_days = (datetime.now(timezone.utc) - timestamp).days
-                            decay_factor = DECAY_WEIGHT + (1 - DECAY_WEIGHT) * math.exp(-age_days / DECAY_SCALE_DAYS)
-                            score = score * decay_factor
+                            age_days = (datetime.now(timezone.utc) - timestamp).total_seconds() / (24 * 60 * 60)
+                            # Use consistent half-life formula: decay = exp(-ln(2) * age / half_life)
+                            ln2 = math.log(2)
+                            decay_factor = math.exp(-ln2 * age_days / DECAY_SCALE_DAYS)
+                            # Apply multiplicative formula: score * ((1 - weight) + weight * decay)
+                            score = score * ((1 - DECAY_WEIGHT) + DECAY_WEIGHT * decay_factor)
                         except (ValueError, TypeError) as e:
                             # Log but continue - timestamp format issue shouldn't break search
                             logger.debug(f"Failed to apply decay for timestamp {payload.get('timestamp')}: {e}")

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "claude-self-reflect",
-  "version": "3.2.3",
+  "version": "3.2.4",
   "description": "Give Claude perfect memory of all your conversations - Installation wizard for Python MCP server",
   "keywords": [
     "claude",
@@ -44,6 +44,7 @@
     "scripts/importer/**/*.py",
     "scripts/delta-metadata-update-safe.py",
     "scripts/force-metadata-recovery.py",
+    "shared/**/*.py",
     ".claude/agents/*.md",
     "config/qdrant-config.yaml",
     "docker-compose.yaml",

package/scripts/import-conversations-unified.py

@@ -25,12 +25,20 @@ sys.path.insert(0, str(scripts_dir))
 from qdrant_client import QdrantClient
 from qdrant_client.models import PointStruct, Distance, VectorParams
 
-# Import the correct normalize_project_name from utils
+# Import normalize_project_name from shared module
+# Add parent directory to path to import shared module
+sys.path.insert(0, str(Path(__file__).parent.parent))
 try:
-    from utils import normalize_project_name
+    from shared.normalization import normalize_project_name
 except ImportError as e:
-    logging.error(f"Failed to import normalize_project_name from utils: {e}")
-    sys.exit(1)
+    logging.error(f"Failed to import normalize_project_name from shared module: {e}")
+    # Fall back to local utils if shared module not found
+    try:
+        from utils import normalize_project_name
+        logging.warning("Using legacy utils.normalize_project_name - consider updating")
+    except ImportError:
+        logging.error("Could not import normalize_project_name from any source")
+        sys.exit(1)
 
 # Set up logging
 logging.basicConfig(

package/scripts/importer/utils/project_normalizer.py

@@ -2,8 +2,17 @@
 
 import hashlib
 import logging
+import sys
 from pathlib import Path
 
+# Import from shared module for consistent normalization
+sys.path.insert(0, str(Path(__file__).parent.parent.parent.parent))
+try:
+    from shared.normalization import normalize_project_name as shared_normalize
+except ImportError:
+    shared_normalize = None
+    logging.warning("Could not import shared normalization module")
+
 logger = logging.getLogger(__name__)
 
 
@@ -20,32 +29,36 @@ class ProjectNormalizer:
         """
         Normalize a project path to a consistent project name.
         
-        CRITICAL: This must match the implementation in utils.py
+        Uses the shared normalization module to ensure consistency
+        across all components.
         
         Examples:
         - "-Users-name-projects-claude-self-reflect" -> "claude-self-reflect"
         - "claude-self-reflect" -> "claude-self-reflect"
         - "/path/to/-Users-name-projects-myapp" -> "myapp"
         """
-        # Get the final component of the path
-        if '/' in project_path:
-            final_component = project_path.split('/')[-1]
-        else:
-            final_component = project_path
+        # Use shared normalization if available
+        if shared_normalize:
+            return shared_normalize(project_path)
+        
+        # Fallback implementation (matches shared module)
+        if not project_path:
+            return ""
+        
+        path = Path(project_path.rstrip('/'))
+        final_component = path.name
         
         # Handle Claude's dash-separated format
         if final_component.startswith('-') and 'projects' in final_component:
-            # Find the last occurrence of 'projects-'
             idx = final_component.rfind('projects-')
             if idx != -1:
-                # Extract everything after 'projects-'
                 project_name = final_component[idx + len('projects-'):]
                 logger.debug(f"Normalized '{project_path}' to '{project_name}'")
                 return project_name
         
         # Already normalized or different format
         logger.debug(f"Project path '{project_path}' already normalized")
-        return final_component
+        return final_component if final_component else path.parent.name
     
     def get_project_name(self, file_path: Path) -> str:
         """

package/shared/__init__.py

@@ -0,0 +1,5 @@
+"""Shared utilities for claude-self-reflect."""
+
+from .normalization import normalize_project_name
+
+__all__ = ['normalize_project_name']

package/shared/normalization.py

@@ -0,0 +1,54 @@
+"""Shared normalization utilities for claude-self-reflect.
+
+This module provides the single source of truth for project name normalization,
+ensuring consistent hashing across import scripts and the MCP server.
+"""
+
+from pathlib import Path
+
+
+def normalize_project_name(project_path: str, _depth: int = 0) -> str:
+    """
+    Normalize project name for consistent hashing across import/search.
+    
+    This is the authoritative normalization function used by both:
+    - Import scripts (import-conversations-unified.py)
+    - MCP server (server.py)
+    
+    Examples:
+        '/Users/name/.claude/projects/-Users-name-projects-myproject' -> 'myproject'
+        '-Users-name-projects-myproject' -> 'myproject'
+        '/path/to/myproject' -> 'myproject'
+        'myproject' -> 'myproject'
+        
+    Special handling for Claude's dash-separated format:
+        When a path component starts with '-' and contains 'projects',
+        we extract everything after 'projects-' as the project name.
+        This handles dashes in project names correctly.
+        
+    Args:
+        project_path: Project path or name in any format
+        _depth: Internal recursion depth counter (for backwards compatibility)
+        
+    Returns:
+        Normalized project name suitable for consistent hashing
+    """
+    if not project_path:
+        return ""
+    
+    path = Path(project_path.rstrip('/'))
+    
+    # Extract the final directory name
+    final_component = path.name
+    
+    # If it's Claude's dash-separated format, extract project name
+    if final_component.startswith('-') and 'projects' in final_component:
+        # Find the last occurrence of 'projects-' to handle edge cases
+        # This correctly extracts 'claude-self-reflect' from:
+        # '-Users-ramakrishnanannaswamy-projects-claude-self-reflect'
+        idx = final_component.rfind('projects-')
+        if idx != -1:
+            return final_component[idx + len('projects-'):]
+    
+    # For regular paths, just return the directory name
+    return final_component if final_component else path.parent.name