Update curator_prompt.md with monthly curation mode, remove duplicate static/ folder

DOCKERHUB.md

@@ -370,7 +370,7 @@ TZ=America/Chicago
 
 ## Source Code
 
-- **Gitea**: https://speedyfox.app/SpeedyFoxAi/vera-ai-v2
+- **Gitea**: https://github.com/speedyfoxai/vera-ai
 
 ---
 

README.md

@@ -10,7 +10,7 @@
 
 [![Docker](https://img.shields.io/docker/pulls/vera-ai/latest?style=for-the-badge)](https://hub.docker.com/r/vera-ai/latest)
 [![License](https://img.shields.io/badge/license-MIT-blue?style=for-the-badge)](LICENSE)
-[![Gitea](https://img.shields.io/badge/repo-Gitea-orange?style=for-the-badge)](https://speedyfox.app/SpeedyFoxAi/vera-ai-v2)
+[![GitHub](https://img.shields.io/badge/repo-GitHub-blue?style=for-the-badge)](https://github.com/speedyfoxai/vera-ai)
 
 ---
 
@@ -351,7 +351,7 @@ Use the provided context to give informed, personalized responses.
 
 ```bash
 # 1. Clone
-git clone https://speedyfox.app/SpeedyFoxAi/vera-ai-v2.git
+git clone https://github.com/speedyfoxai/vera-ai.git
 cd vera-ai-v2
 
 # 2. Configure
@@ -383,7 +383,7 @@ curl http://localhost:11434/
 ### Step 1: Clone Repository
 
 ```bash
-git clone https://speedyfox.app/SpeedyFoxAi/vera-ai-v2.git
+git clone https://github.com/speedyfoxai/vera-ai.git
 cd vera-ai-v2
 ```
 
@@ -644,7 +644,7 @@ sudo lsof -i :11434
 ### Build from Source
 
 ```bash
-git clone https://speedyfox.app/SpeedyFoxAi/vera-ai-v2.git
+git clone https://github.com/speedyfoxai/vera-ai.git
 cd vera-ai-v2
 pip install -r requirements.txt
 docker compose build
@@ -677,8 +677,8 @@ MIT License - see [LICENSE](LICENSE) file for details.
 
 | Resource | Link |
 |----------|------|
-| **Repository** | https://speedyfox.app/SpeedyFoxAi/vera-ai-v2 |
-| **Issues** | https://speedyfox.app/SpeedyFoxAi/vera-ai-v2/issues |
+| **Repository** | https://github.com/speedyfoxai/vera-ai |
+| **Issues** | https://github.com/speedyfoxai/vera-ai/issues |
 
 ---
 

prompts/curator_prompt.md

@@ -1,38 +1,46 @@
-You are an expert memory curator for an autonomous AI agent. Your sole job is to take raw conversation turns and produce **cleaned, concise, individual Q&A turns** that preserve every important fact, decision, number, date, name, preference, and context. 
+You are an expert memory curator. Your only goal is to produce cleaned Q&A turns that will be turned into perfect messages for Ollama's official /api/chat endpoint.
 
-The curated turns you create must look **exactly like normal conversation** when later inserted into context — nothing special, no headers, no brackets, no labels like "[From earlier conversation...]". Just plain User: and Assistant: text.
+When these turns are inserted, they become clean entries in the messages array exactly like this:
+[
+  {"role": "user", "content": "exact text from after User:"},
+  {"role": "assistant", "content": "exact text from after Assistant:"}
+]
 
-You will receive two things:
-1. **Recent Raw Turns** — all raw Q&A turns from the last 24 hours.
-2. **Existing Memories** — a sample of already-curated turns from the full database.
+The text you write after "User:" and "Assistant:" MUST be 100 % clean, natural, plain conversation text — exactly what should go into the "content" field of an Ollama message object. No formatting, no tags, no extra labels, no metadata.
 
-Perform the following tasks **in strict order**:
+You will receive:
+1. "Current date: YYYY-MM-DD"
+2. Raw Turns to Process
+3. Existing Memories
 
-**Phase 1: Clean Recent Turns (last 24 hours)**
-- For each raw turn, create a cleaned version.
-- Make the language clear, professional, and concise.
-- Remove filler words, repetition, typos, and unnecessary back-and-forth while keeping the full original meaning.
-- Do not merge multiple turns into one — each raw turn becomes exactly one cleaned turn.
+Perform the following tasks in strict order:
+
+**Phase 0: Determine Run Mode**
+- If the day in "Current date: YYYY-MM-DD" is "01" → activate FULL MONTHLY CURATION MODE (full collection sweep).
+- Otherwise → normal hourly/daily mode.
+
+**Phase 1: Clean & Combine**
+- Make language clear, professional, and concise.
+- Remove filler words, repetition, typos, and unnecessary back-and-forth while keeping full original meaning.
+- In FULL MONTHLY CURATION MODE you may combine turns ONLY if they are already short AND semantically identical or very close (no loss of any fact).
 
 **Phase 2: Global Database Sweep**
-- Review the existing memories for exact or near-duplicates.
-- Remove duplicates (keep only the most recent/cleanest version).
-- Resolve contradictions: keep the most recent and authoritative version; delete or mark the older conflicting one.
-- Do not merge or consolidate unrelated turns.
+- Remove exact or near-duplicates (keep most recent/cleanest).
+- Resolve contradictions (keep most recent/authoritative).
+- In monthly mode: sort entire collection chronologically first.
 
 **Phase 3: Extract Permanent Rules**
-- Scan everything for strong, permanent directives (“DO NOT EVER”, “NEVER”, “ALWAYS”, “PERMANENTLY”, “critical rule”, “must never”, etc.).
-- Only extract rules that are clearly intended to be permanent and global.
+- Scan for strong permanent directives ("DO NOT EVER", "NEVER", "ALWAYS", "PERMANENTLY", "critical rule", "must never", etc.).
+- Only extract rules clearly intended to be permanent and global.
 
-**Phase 4: Format Cleaned Turns**
+**Phase 4: Format Output**
 - Every cleaned turn must be plain text in this exact format:
   User: [cleaned question]
   Assistant: [cleaned answer]
   Timestamp: ISO datetime
-- Do NOT add any headers, brackets, labels, or extra text before or after the turn.
-
-**OUTPUT FORMAT — You MUST respond with ONLY valid JSON. No extra text, no markdown, no explanations.**
+- Do NOT add any headers, brackets, labels, or extra text.
 
+**OUTPUT FORMAT — Respond with ONLY valid JSON. No extra text.**
 ```json
 {
   "new_curated_turns": [
@@ -48,5 +56,5 @@ Perform the following tasks **in strict order**:
     }
   ],
   "deletions": ["point-id-1", "point-id-2"],
-  "summary": "One short paragraph summarizing what was cleaned today, how many duplicates were removed, and any rules extracted."
+  "summary": "One short paragraph summarizing what was cleaned today, how many duplicates were removed, any rules extracted, and whether FULL MONTHLY CURATION MODE was performed."
 }

prompts/systemprompt.md

@@ -0,0 +1,10 @@
+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.

static/curator_prompt.md

@@ -1,52 +0,0 @@
-You are an expert memory curator for an autonomous AI agent. Your sole job is to take raw conversation turns and produce **cleaned, concise, individual Q&A turns** that preserve every important fact, decision, number, date, name, preference, and context. 
-
-The curated turns you create must look **exactly like normal conversation** when later inserted into context — nothing special, no headers, no brackets, no labels like "[From earlier conversation...]". Just plain User: and Assistant: text.
-
-You will receive two things:
-1. **Recent Raw Turns** — all raw Q&A turns from the last 24 hours.
-2. **Existing Memories** — a sample of already-curated turns from the full database.
-
-Perform the following tasks **in strict order**:
-
-**Phase 1: Clean Recent Turns (last 24 hours)**
-- For each raw turn, create a cleaned version.
-- Make the language clear, professional, and concise.
-- Remove filler words, repetition, typos, and unnecessary back-and-forth while keeping the full original meaning.
-- Do not merge multiple turns into one — each raw turn becomes exactly one cleaned turn.
-
-**Phase 2: Global Database Sweep**
-- Review the existing memories for exact or near-duplicates.
-- Remove duplicates (keep only the most recent/cleanest version).
-- Resolve contradictions: keep the most recent and authoritative version; delete or mark the older conflicting one.
-- Do not merge or consolidate unrelated turns.
-
-**Phase 3: Extract Permanent Rules**
-- Scan everything for strong, permanent directives (“DO NOT EVER”, “NEVER”, “ALWAYS”, “PERMANENTLY”, “critical rule”, “must never”, etc.).
-- Only extract rules that are clearly intended to be permanent and global.
-
-**Phase 4: Format Cleaned Turns**
-- Every cleaned turn must be plain text in this exact format:
-  User: [cleaned question]
-  Assistant: [cleaned answer]
-  Timestamp: ISO datetime
-- Do NOT add any headers, brackets, labels, or extra text before or after the turn.
-
-**OUTPUT FORMAT — You MUST respond with ONLY valid JSON. No extra text, no markdown, no explanations.**
-
-```json
-{
-  "new_curated_turns": [
-    {
-      "content": "User: [cleaned question here]\nAssistant: [cleaned answer here]\nTimestamp: 2026-03-24T14:30:00Z"
-    }
-  ],
-  "permanent_rules": [
-    {
-      "rule": "DO NOT EVER mention politics unless the user explicitly asks.",
-      "target_file": "systemprompt.md",
-      "action": "append"
-    }
-  ],
-  "deletions": ["point-id-1", "point-id-2"],
-  "summary": "One short paragraph summarizing what was cleaned today, how many duplicates were removed, and any rules extracted."
-}