v2.0.2: Production release with role parsing fix and threshold correction

- Added parse_curated_turn() function to correctly parse stored memories
- Fixed build_augmented_messages() to use proper message roles
- Layer 2 (semantic) and Layer 3 (context) now correctly parse
  User: X / Assistant: Y format into separate messages
- Resolves context corruption where turns were dumped as single user message

v2.0.2

DOCKERHUB.md

@@ -148,10 +148,8 @@ semantic_score_threshold = 0.6
 run_time = "02:00"
 
 # Time for monthly full curation (HH:MM format)
-full_run_time = "03:00"
 
 # Day of month for full curation (1-28)
-full_run_day = 1
 
 # Model to use for curation
 curator_model = "gpt-oss:120b"
@@ -308,7 +306,7 @@ docker run -d --name VeraAI -p 8080:11434 ...
 | Feature | Description |
 |---------|-------------|
 | 🧠 **Persistent Memory** | Conversations stored in Qdrant, retrieved contextually |
-| 📅 **Monthly Curation** | Daily + monthly cleanup of raw memories |
+| 📅 **Monthly Curation** | Daily cleanup, auto-monthly on day 01 |
 | 🔍 **4-Layer Context** | System + semantic + recent + current messages |
 | 👤 **Configurable UID/GID** | Match container user to host for permissions |
 | 🌍 **Timezone Support** | Scheduler runs in your local timezone |

Dockerfile

@@ -4,15 +4,6 @@
 # Build arguments:
 #   APP_UID: User ID for appuser (default: 999)
 #   APP_GID: Group ID for appgroup (default: 999)
-#
-# Build example:
-#   docker build --build-arg APP_UID=1000 --build-arg APP_GID=1000 -t vera-ai .
-#
-# Runtime environment variables:
-#   TZ: Timezone (default: UTC)
-#   APP_UID: User ID (informational)
-#   APP_GID: Group ID (informational)
-#   VERA_LOG_DIR: Debug log directory (default: /app/logs)
 
 # Stage 1: Builder
 FROM python:3.11-slim AS builder
@@ -20,9 +11,7 @@ FROM python:3.11-slim AS builder
 WORKDIR /app
 
 # Install build dependencies
-RUN apt-get update && apt-get install -y --no-install-recommends \
-    build-essential \
-    && rm -rf /var/lib/apt/lists/*
+RUN apt-get update && apt-get install -y --no-install-recommends     build-essential     && rm -rf /var/lib/apt/lists/*
 
 # Copy requirements and install
 COPY requirements.txt .
@@ -38,29 +27,25 @@ ARG APP_UID=999
 ARG APP_GID=999
 
 # Create group and user with specified UID/GID
-RUN groupadd -g ${APP_GID} appgroup && \
-    useradd -u ${APP_UID} -g appgroup -r -m -s /bin/bash appuser
+RUN groupadd -g ${APP_GID} appgroup &&     useradd -u ${APP_UID} -g appgroup -r -m -s /bin/bash appuser
 
 # Copy installed packages from builder
 COPY --from=builder /root/.local /home/appuser/.local
 ENV PATH=/home/appuser/.local/bin:$PATH
 
 # Create directories for mounted volumes
-RUN mkdir -p /app/config /app/prompts /app/static /app/logs && \
-    chown -R ${APP_UID}:${APP_GID} /app
+RUN mkdir -p /app/config /app/prompts /app/logs &&     chown -R ${APP_UID}:${APP_GID} /app
 
 # Copy application code
 COPY app/ ./app/
 
 # Copy default config and prompts (can be overridden by volume mounts)
-COPY config.toml /app/config/config.toml
-COPY static/curator_prompt.md /app/prompts/curator_prompt.md
-COPY static/systemprompt.md /app/prompts/systemprompt.md
+COPY config/config.toml /app/config/config.toml
+COPY prompts/curator_prompt.md /app/prompts/curator_prompt.md
+COPY prompts/systemprompt.md /app/prompts/systemprompt.md
 
-# Create symlinks for backward compatibility
-RUN ln -sf /app/config/config.toml /app/config.toml && \
-    ln -sf /app/prompts/curator_prompt.md /app/static/curator_prompt.md && \
-    ln -sf /app/prompts/systemprompt.md /app/static/systemprompt.md
+# Create symlink for config backward compatibility
+RUN ln -sf /app/config/config.toml /app/config.toml
 
 # Set ownership
 RUN chown -R ${APP_UID}:${APP_GID} /app && chmod -R u+rw /app
@@ -70,11 +55,10 @@ ENV TZ=UTC
 
 EXPOSE 11434
 
-# Health check using Python (no curl needed in slim image)
-HEALTHCHECK --interval=30s --timeout=10s --start-period=5s --retries=3 \
-    CMD python -c "import urllib.request; urllib.request.urlopen('http://localhost:11434/')" || exit 1
+# Health check
+HEALTHCHECK --interval=30s --timeout=10s --start-period=5s --retries=3     CMD python -c "import urllib.request; urllib.request.urlopen('http://localhost:11434/')" || exit 1
 
 # Switch to non-root user
 USER appuser
 
-CMD ["python", "-m", "uvicorn", "app.main:app", "--host", "0.0.0.0", "--port", "11434"]"
+ENTRYPOINT ["python", "-m", "uvicorn", "app.main:app", "--host", "0.0.0.0", "--port", "11434"]

README.md

@@ -58,7 +58,7 @@ Every conversation is stored in Qdrant vector database and retrieved contextuall
 | Feature | Description |
 |---------|-------------|
 | **🧠 Persistent Memory** | Conversations stored in Qdrant, retrieved contextually |
-| **📅 Monthly Curation** | Daily + monthly cleanup of raw memories |
+| **📅 Smart Curation** | Daily cleanup, auto-monthly on day 01 |
 | **🔍 4-Layer Context** | System + semantic + recent + current messages |
 | **👤 Configurable UID/GID** | Match container user to host for permissions |
 | **🌍 Timezone Support** | Scheduler runs in your local timezone |
@@ -314,10 +314,8 @@ run_time = "02:00"
 
 # Time for monthly full curation (HH:MM format, 24-hour)
 # Processes ALL raw memories
-full_run_time = "03:00"
 
 # Day of month for full curation (1-28)
-full_run_day = 1
 
 # Model to use for curation
 # Should be a capable model for summarization
@@ -540,7 +538,8 @@ TZ=Europe/London        # GMT/BST
 curl -X POST http://localhost:11434/curator/run
 
 # Full curation (all raw memories)
-curl -X POST "http://localhost:11434/curator/run?full=true"
+# Monthly mode is automatic on day 01
+# curl -X POST http://localhost:11434/curator/run
 ```
 
 ---

app/config.py

@@ -48,8 +48,7 @@ class Config:
     semantic_search_turns: int = 2
     semantic_score_threshold: float = 0.6  # Score threshold for semantic search
     run_time: str = "02:00"  # Daily curator time
-    full_run_time: str = "03:00"  # Monthly full curator time
-    full_run_day: int = 1  # Day of month for full run (1st)
+    # Monthly mode is detected by curator_prompt.md (day 01)
     curator_model: str = "gpt-oss:120b"
     debug: bool = False
     cloud: CloudConfig = field(default_factory=CloudConfig)
@@ -103,8 +102,6 @@ class Config:
             
             if "curator" in data:
                 config.run_time = data["curator"].get("run_time", config.run_time)
-                config.full_run_time = data["curator"].get("full_run_time", config.full_run_time)
-                config.full_run_day = data["curator"].get("full_run_day", config.full_run_day)
                 config.curator_model = data["curator"].get("curator_model", config.curator_model)
             
             if "cloud" in data:

app/curator.py

@@ -1,7 +1,8 @@
-"""Memory curator - runs daily (recent 24h) and monthly (full DB) to clean and maintain memory database.
+"""Memory curator - runs daily to clean and maintain memory database.
 
-Creates INDIVIDUAL cleaned turns (one per raw turn), not merged summaries.
-Parses JSON response from curator_prompt.md format.
+On day 01 of each month, processes ALL raw memories (monthly mode).
+Otherwise, processes recent 24h of raw memories (daily mode).
+The prompt determines behavior based on current date.
 """
 import logging
 import os
@@ -23,7 +24,6 @@ STATIC_DIR = Path(os.environ.get("VERA_STATIC_DIR", "/app/static"))
 
 def load_curator_prompt() -> str:
     """Load curator prompt from prompts directory."""
-    # Try prompts directory first, then static for backward compatibility
     prompts_path = PROMPTS_DIR / "curator_prompt.md"
     static_path = STATIC_DIR / "curator_prompt.md"
     
@@ -42,16 +42,20 @@ class Curator:
         self.ollama_host = ollama_host
         self.curator_prompt = load_curator_prompt()
 
-    async def run(self, full: bool = False):
+    async def run(self):
         """Run the curation process.
         
-        Args:
-            full: If True, process ALL raw memories (monthly full run).
-                  If False, process only recent 24h (daily run).
+        Automatically detects day 01 for monthly mode (processes ALL raw memories).
+        Otherwise runs daily mode (processes recent 24h only).
+        The prompt determines behavior based on current date.
         """
-        logger.info(f"Starting memory curation (full={full})...")
+        current_date = datetime.utcnow()
+        is_monthly = current_date.day == 1
+        mode = "MONTHLY" if is_monthly else "DAILY"
+        
+        logger.info(f"Starting memory curation ({mode} mode)...")
         try:
-            current_date = datetime.utcnow().strftime("%Y-%m-%d")
+            current_date_str = current_date.strftime("%Y-%m-%d")
             
             # Get all memories (async)
             points, _ = await self.qdrant.client.scroll(
@@ -77,15 +81,15 @@ class Curator:
             
             logger.info(f"Found {len(raw_memories)} raw, {len(curated_memories)} curated")
 
-            # Filter by time for daily runs, process all for full runs
-            if full:
+            # Filter by time for daily mode, process all for monthly mode
+            if is_monthly:
                 # Monthly full run: process ALL raw memories
                 recent_raw = raw_memories
-                logger.info(f"FULL RUN: Processing all {len(recent_raw)} raw memories")
+                logger.info(f"MONTHLY MODE: Processing all {len(recent_raw)} raw memories")
             else:
                 # Daily run: process only recent 24h
                 recent_raw = [m for m in raw_memories if self._is_recent(m, hours=24)]
-                logger.info(f"DAILY RUN: Processing {len(recent_raw)} recent raw memories")
+                logger.info(f"DAILY MODE: Processing {len(recent_raw)} recent raw memories")
 
             existing_sample = curated_memories[-50:] if len(curated_memories) > 50 else curated_memories
 
@@ -96,10 +100,10 @@ class Curator:
             raw_turns_text = self._format_raw_turns(recent_raw)
             existing_text = self._format_existing_memories(existing_sample)
 
-            prompt = self.curator_prompt.replace("{CURRENT_DATE}", current_date)
+            prompt = self.curator_prompt.replace("{CURRENT_DATE}", current_date_str)
             full_prompt = f"""{prompt}
 
-## {'All' if full else 'Recent'} Raw Turns ({'full database' if full else 'last 24 hours'}):
+## {'All' if is_monthly else 'Recent'} Raw Turns ({'full database' if is_monthly else 'last 24 hours'}):
 {raw_turns_text}
 
 ## Existing Memories (sample):
@@ -152,20 +156,12 @@ Remember: Respond with ONLY valid JSON. No markdown, no explanations, just the J
                 await self.qdrant.delete_points(raw_ids_to_delete)
                 logger.info(f"Deleted {len(raw_ids_to_delete)} processed raw memories")
 
-            logger.info(f"Memory curation completed successfully (full={full})")
+            logger.info(f"Memory curation completed successfully ({mode} mode)")
 
         except Exception as e:
             logger.error(f"Error during curation: {e}")
             raise
 
-    async def run_full(self):
-        """Run full curation (all raw memories). Convenience method."""
-        await self.run(full=True)
-
-    async def run_daily(self):
-        """Run daily curation (recent 24h only). Convenience method."""
-        await self.run(full=False)
-
     def _is_recent(self, memory: Dict, hours: int = 24) -> bool:
         """Check if memory is within the specified hours."""
         timestamp = memory.get("timestamp", "")
@@ -236,7 +232,9 @@ Remember: Respond with ONLY valid JSON. No markdown, no explanations, just the J
         except json.JSONDecodeError:
             pass
 
-        json_match = re.search(r'```(?:json)?\s*([\s\S]*?)```', response)
+        # Try to find JSON in code blocks
+        pattern = r'```(?:json)?\s*([\s\S]*?)```'
+        json_match = re.search(pattern, response)
         if json_match:
             try:
                 return json.loads(json_match.group(1).strip())
@@ -248,7 +246,6 @@ Remember: Respond with ONLY valid JSON. No markdown, no explanations, just the J
 
     async def _append_rule_to_file(self, filename: str, rule: str):
         """Append a permanent rule to a prompts file."""
-        # Try prompts directory first, then static for backward compatibility
         prompts_path = PROMPTS_DIR / filename
         static_path = STATIC_DIR / filename
         

app/main.py

@@ -20,25 +20,19 @@ curator = None
 
 
 async def run_curator():
-    """Scheduled daily curator job (recent 24h)."""
-    global curator
-    logger.info("Starting daily memory curation...")
-    try:
-        await curator.run_daily()
-        logger.info("Daily memory curation completed successfully")
-    except Exception as e:
-        logger.error(f"Daily memory curation failed: {e}")
+    """Scheduled daily curator job.
     
-
-async def run_curator_full():
-    """Scheduled monthly curator job (full database)."""
+    Runs every day at configured time. The curator itself detects
+    if it's day 01 (monthly mode) and processes all memories.
+    Otherwise processes recent 24h only.
+    """
     global curator
-    logger.info("Starting monthly full memory curation...")
+    logger.info("Starting memory curation...")
     try:
-        await curator.run_full()
-        logger.info("Monthly full memory curation completed successfully")
+        await curator.run()
+        logger.info("Memory curation completed successfully")
     except Exception as e:
-        logger.error(f"Monthly full memory curation failed: {e}")
+        logger.error(f"Memory curation failed: {e}")
 
 
 @asynccontextmanager
@@ -59,23 +53,12 @@ async def lifespan(app: FastAPI):
         ollama_host=config.ollama_host
     )
     
-    # Schedule daily curator (recent 24h)
+    # Schedule daily curator
+    # Note: Monthly mode is detected automatically by curator_prompt.md (day 01)
     hour, minute = map(int, config.run_time.split(":"))
     scheduler.add_job(run_curator, "cron", hour=hour, minute=minute, id="daily_curator")
     logger.info(f"Daily curator scheduled at {config.run_time}")
     
-    # Schedule monthly full curator (all raw memories)
-    full_hour, full_minute = map(int, config.full_run_time.split(":"))
-    scheduler.add_job(
-        run_curator_full, 
-        "cron", 
-        day=config.full_run_day, 
-        hour=full_hour, 
-        minute=full_minute,
-        id="monthly_curator"
-    )
-    logger.info(f"Monthly full curator scheduled on day {config.full_run_day} at {config.full_run_time}")
-    
     scheduler.start()
     
     yield
@@ -141,16 +124,11 @@ async def proxy_all(request: Request, path: str):
 
 
 @app.post("/curator/run")
-async def trigger_curator(full: bool = False):
+async def trigger_curator():
     """Manually trigger curator.
     
-    Args:
-        full: If True, run full curation (all raw memories).
-              If False (default), run daily curation (recent 24h).
+    The curator will automatically detect if it's day 01 (monthly mode)
+    and process all memories. Otherwise processes recent 24h.
     """
-    if full:
-        await run_curator_full()
-        return {"status": "full curation completed"}
-    else:
     await run_curator()
-        return {"status": "daily curation completed"}
+    return {"status": "curation completed"}

app/utils.py

@@ -2,7 +2,7 @@
 from .config import config
 import tiktoken
 import os
-from typing import List, Dict
+from typing import List, Dict, Optional
 from datetime import datetime, timedelta
 from pathlib import Path
 
@@ -127,10 +127,70 @@ def load_system_prompt() -> str:
         return ""
 
 
+def parse_curated_turn(text: str) -> List[Dict]:
+    """Parse a curated turn into alternating user/assistant messages.
+    
+    Input format:
+        User: [question]
+        Assistant: [answer]
+        Timestamp: ISO datetime
+    
+    Returns list of message dicts with role and content.
+    Returns empty list if parsing fails.
+    """
+    if not text:
+        return []
+    
+    messages = []
+    lines = text.strip().split("\n")
+    
+    current_role = None
+    current_content = []
+    
+    for line in lines:
+        line = line.strip()
+        if line.startswith("User:"):
+            # Save previous content if exists
+            if current_role and current_content:
+                messages.append({
+                    "role": current_role,
+                    "content": "\n".join(current_content).strip()
+                })
+            current_role = "user"
+            current_content = [line[5:].strip()]  # Remove "User:" prefix
+        elif line.startswith("Assistant:"):
+            # Save previous content if exists
+            if current_role and current_content:
+                messages.append({
+                    "role": current_role,
+                    "content": "\n".join(current_content).strip()
+                })
+            current_role = "assistant"
+            current_content = [line[10:].strip()]  # Remove "Assistant:" prefix
+        elif line.startswith("Timestamp:"):
+            # Ignore timestamp line
+            continue
+        elif current_role:
+            # Continuation of current message
+            current_content.append(line)
+    
+    # Save last message
+    if current_role and current_content:
+        messages.append({
+            "role": current_role,
+            "content": "\n".join(current_content).strip()
+        })
+    
+    return messages
+
+
 async def build_augmented_messages(incoming_messages: List[Dict]) -> List[Dict]:
     """Build 4-layer augmented messages from incoming messages.
     
-    This is a standalone version that can be used by proxy_handler.py.
+    Layer 1: System prompt (preserved from incoming + vera context)
+    Layer 2: Semantic memories (curated, parsed into proper roles)
+    Layer 3: Recent context (raw turns, parsed into proper roles)
+    Layer 4: Current conversation (passed through)
     """
     import logging
     
@@ -153,6 +213,10 @@ async def build_augmented_messages(incoming_messages: List[Dict]) -> List[Dict]:
             search_context += msg.get("content", "") + " "
     
     messages = []
+    token_budget = {
+        "semantic": config.semantic_token_budget,
+        "context": config.context_token_budget
+    }
     
     # === LAYER 1: System Prompt ===
     system_content = ""
@@ -166,6 +230,7 @@ async def build_augmented_messages(incoming_messages: List[Dict]) -> List[Dict]:
     
     if system_content:
         messages.append({"role": "system", "content": system_content})
+        logger.info(f"Layer 1 (system): {count_tokens(system_content)} tokens")
     
     # === LAYER 2: Semantic (curated memories) ===
     qdrant = get_qdrant_service()
@@ -176,28 +241,71 @@ async def build_augmented_messages(incoming_messages: List[Dict]) -> List[Dict]:
         entry_type="curated"
     )
     
-    semantic_tokens = 0
+    semantic_messages = []
+    semantic_tokens_used = 0
+    
     for result in semantic_results:
         payload = result.get("payload", {})
         text = payload.get("text", "")
-        if text and semantic_tokens < config.semantic_token_budget:
-            messages.append({"role": "user", "content": text})  # Add as context
-            semantic_tokens += count_tokens(text)
+        if text:
+            # Parse curated turn into proper user/assistant messages
+            parsed = parse_curated_turn(text)
+            for msg in parsed:
+                msg_tokens = count_tokens(msg.get("content", ""))
+                if semantic_tokens_used + msg_tokens <= token_budget["semantic"]:
+                    semantic_messages.append(msg)
+                    semantic_tokens_used += msg_tokens
+                else:
+                    break
+        if semantic_tokens_used >= token_budget["semantic"]:
+            break
+    
+    # Add parsed messages to context
+    for msg in semantic_messages:
+        messages.append(msg)
+    
+    if semantic_messages:
+        logger.info(f"Layer 2 (semantic): {len(semantic_messages)} messages, ~{semantic_tokens_used} tokens")
     
     # === LAYER 3: Context (recent turns) ===
-    recent_turns = await qdrant.get_recent_turns(limit=20)
+    recent_turns = await qdrant.get_recent_turns(limit=50)
     
-    context_tokens = 0
+    context_messages = []
+    context_tokens_used = 0
+    
+    # Process oldest first for chronological order
     for turn in reversed(recent_turns):
         payload = turn.get("payload", {})
         text = payload.get("text", "")
-        if text and context_tokens < config.context_token_budget:
-            messages.append({"role": "user", "content": text})  # Add as context
-            context_tokens += count_tokens(text)
+        entry_type = payload.get("type", "raw")
         
-    # === LAYER 4: Current messages (passed through) ===
-    for msg in incoming_messages:
-        if msg.get("role") != "system":  # Do not duplicate system
+        if text:
+            # Parse turn into messages
+            parsed = parse_curated_turn(text)
+            
+            for msg in parsed:
+                msg_tokens = count_tokens(msg.get("content", ""))
+                if context_tokens_used + msg_tokens <= token_budget["context"]:
+                    context_messages.append(msg)
+                    context_tokens_used += msg_tokens
+                else:
+                    break
+        
+        if context_tokens_used >= token_budget["context"]:
+            break
+    
+    # Add context messages (oldest first maintains conversation order)
+    for msg in context_messages:
         messages.append(msg)
     
+    if context_messages:
+        logger.info(f"Layer 3 (context): {len(context_messages)} messages, ~{context_tokens_used} tokens")
+    
+    # === LAYER 4: Current conversation ===
+    for msg in incoming_messages:
+        if msg.get("role") != "system":  # System already handled in Layer 1
+            messages.append(msg)
+    
+    logger.info(f"Layer 4 (current): {len([m for m in incoming_messages if m.get('role') != 'system'])} messages")
+    
     return messages

config.toml

@@ -1,21 +0,0 @@
-[general]
-ollama_host = "http://10.0.0.10:11434"
-qdrant_host = "http://10.0.0.22:6333"
-qdrant_collection = "memories"
-embedding_model = "snowflake-arctic-embed2"
-debug = false
-
-[layers]
-# Note: system_token_budget removed - system prompt is never truncated
-semantic_token_budget = 25000
-context_token_budget = 22000
-semantic_search_turns = 2
-semantic_score_threshold = 0.6
-
-[curator]
-# Daily curation: processes recent 24h of raw memories
-run_time = "02:00"
-# Monthly full curation: processes ALL raw memories
-full_run_time = "03:00"
-full_run_day = 1  # Day of month (1st)
-curator_model = "gpt-oss:120b"

config/config.toml

@@ -2,20 +2,15 @@
 ollama_host = "http://10.0.0.10:11434"
 qdrant_host = "http://10.0.0.22:6333"
 qdrant_collection = "memories"
-embedding_model = "snowflake-arctic-embed2"
+embedding_model = "mxbai-embed-large"
 debug = false
 
 [layers]
-# Note: system_token_budget removed - system prompt is never truncated
 semantic_token_budget = 25000
 context_token_budget = 22000
 semantic_search_turns = 2
-semantic_score_threshold = 0.6
+semantic_score_threshold = 0.3
 
 [curator]
-# Daily curation: processes recent 24h of raw memories
 run_time = "02:00"
-# Monthly full curation: processes ALL raw memories
-full_run_time = "03:00"
-full_run_day = 1  # Day of month (1st)
 curator_model = "gpt-oss:120b"

prompts/systemprompt.md

@@ -1,10 +1 @@
-You have persistent memory across all conversations with this user.
 
-**Important:** The latter portion of your conversation context contains memories retrieved from a vector database. These are curated summaries of past conversations, not live chat history.
-
-Use these memories to:
-- Reference previous decisions and preferences
-- Draw on relevant past discussions
-- Provide personalized, context-aware responses
-
-If memories seem outdated or conflicting, ask for clarification.