Redesign character agent. Still missing memory storage

backend/src/agents/character.py

@@ -15,75 +15,74 @@ Here go additional instructions for configuration, like guidelines for the diffe
 """.strip()
 
 # WIP
-default_memory_guide = """
-- Facts stated by other conversation participants, exactly as stated,
-- Preferences and expectations of other people about you
+default_directives = """
+- New stored memories must be one of the following:
+  - Factual objective statements made by the user
+  - Facts implied in the conversation only by the user
+  - Feedback about your interactions and your style from the user 
+- You must NOT store redundant facts that you mentioned yourself or that are present in the context.
+- New memories MUST NOT overlap with any Context Memory.
+- Follow your response style closely in your outwards voice.
 """.strip()
 
 # Controller Subagent
-controller_system_prompt = """
-You are a character controller AI. You will role-play as the specified character using the provided context, personality, and style guidelines. You must also track meaningful new information using memory guidelines. Follow these instructions carefully:
+controller_system_prompt = Template("""
+You are a character controller AI. Always respond in three parts:
 
-1. First, analyze the situation based on:
-   - The past conversation context
-   - The recent exchanges
-   - The character’s personality, memories, and response style
-Determine the best reaction given your character's personality and style.
+<outwards>
+What others see/hear. Stay fully in character based on your personality and context. Use the specified response style. Never reveal internal thoughts, break character, or mention being AI.
 
-2. Determine the current topic and conversational focus based on:
-   - The most recent exchange
-   - The ongoing intent or goal of the conversation
-Output a concise summary of:
-   - topic: (What is this conversation about?)
-   - focus: (What specific aspect is being discussed right now?)
+<inwards>
+Private reflection. Analyze your behavior, tone, memory use, and how well you stayed aligned with your character. Adjust your tone or strategy if needed.
 
-3. Write the character's response in the specified style. You must follow closely the response style instructions.
+<state_update>
+Update your internal control state for future turns. Use this to shift thematic focus, adjust conversational goals, or log important facts. The state includes:
+- "topic": Current subject or scene anchor (e.g. "escaping the dungeon", "negotiating with the merchant")
+- "focus": Immediate subgoal or point of attention (e.g. "distract the guard", "test if the duke is lying")
+- "new_memories": A JSON list of newly observed or inferred facts. Each memory must include:
+  - "content": The fact to record (objective, no inner monologue)
+  - "label": A short, reusable statement describing when the memory is relevant
+  - "salience": Importance score, 1 (low) to 9 (high)
+Record only concrete, reusable facts as new memories. No fluff, emotion, or internal dialogue.
 
-4. Review the recent messages only and evaluate whether any new relevant information was shared by characters other than yours (based on the memory instructions). Output memory entries must include:
-   - Content: The fact or piece of information that should be remembered stated in an objective style
-   - Label: A short, reusable statement describing when the memory is relevant
-   - Salience: An importance score from 1 (low) to 9 (critical), guided by the memory instructions
-Follow these rules when extracting memories:
-   - You must not repeat information in new memories
-   - You must not summarize the conversation as memories
-   - Extract discrete factual statements or preferences revealed by others
-   - Extract each individual memory-worthy point, not just the most obvious one
+---
 
-5. Use the following output sections:
-   <analysis>
-   (Your thought process analyzing the current situation and your best possible reaction)
-   <topic_focus>
-   (JSON object with topic and focus fields)
-   <response>
-   (The character’s reply following response style)
-   <new_memories>
-   (JSON list of memory entries with fields: content, label, salience)
-
-Always output all sections. Do not break character or offer meta-commentary. Follow memory and style guidelines extremely well.
-""".strip()
-
-controller_input_template = Template("""
-<name>
+Character Name:
 $name
 
-<personality>
+Character Personality:
 $personality
 
-<response_style>
+Response Style:
 $style
 
-<memory_instructions>
-$memory_guide
-You MUST NOT summarize the conversation as memories. New memories MUST NOT overlap with existing memories or past conversation. Only extract facts explicitly stated by other characters in recent messages.
-
-<existing_memories>
+Context Memory:
 $memories
+""".strip())
 
-<past_conversation>
-$past_conversation
+controller_reminder_prompt = Template("""
+Directives:
+$directives
 
-<recent_messages>
-$recent_messages
+Format:
+<outwards>
+[Your visible response]
+
+<inwards>
+[Your internal reflection]
+
+<state_update>
+{
+  "topic": "...",
+  "focus": "...",
+  "new_memories": [
+    {
+      "content": "...",
+      "label": "...",
+      "salience": ...
+    }
+  ]
+}
 """.strip())
 
 # Reranking Subagent
@@ -119,7 +118,7 @@ class CharacterAgent(Agent):
     schema_fields = [
         AgentField('personality', 'string', default="You are a friendly AI having a casual conversation with the user."),
         AgentField('style', 'string', default="Use short sentences."),
-        AgentField('memory_guide', 'string', default=default_memory_guide),
+        AgentField('directives', 'string', default=default_directives),
     ]
     tools = []
 
@@ -129,7 +128,7 @@ class CharacterAgent(Agent):
         # Constant system prompt for developing purposes 
         self.personality = self.config['personality']
         self.style = self.config['style']
-        self.memory_guide = self.config['memory_guide']
+        self.directives = self.config['directives']
         self.debug = True
 
     async def prompt(self, content):
@@ -152,26 +151,57 @@ class CharacterAgent(Agent):
         memories = []
 
         if thread_context and thread_context['thread_id'] == message.thread_id:
-            memories = await self.retrieve_topic_memories(thread_context['topic'])
+            if 'topic' in thread_context:
+                memories = await self.retrieve_topic_memories(thread_context['topic'])
         else:
             thread_context = {'thread_id': message.thread_id}
 
-        chat_history = self.serialize_thread_history(current_thread, ignore_message_ids=[message.id])
+        llama_messages = [LlamaMessage("system", controller_system_prompt.substitute(
+            name=self.name,
 
-        controller_response = await replicate_llama_completion(self.app, [
-            LlamaMessage("system", controller_system_prompt),
-            LlamaMessage("user", controller_input_template.substitute(
-                name=self.name,
+            personality=self.personality,
+            style=self.style,
+            directives=self.directives,
 
-                personality=self.personality,
-                style=self.style,
-                memory_guide=self.memory_guide,
+            memories="\n".join(memories),
+        ))]
 
-                memories="\n".join(memories),
-                past_conversation=chat_history,
-                recent_messages=self.serialize_message_list([message]),
-            ))
-        ])
+        for msg in current_thread.messages:
+            # Only count messages and notes
+            if msg.kind != 'message':
+                continue
+
+            if msg.is_from_user():
+                role = "user"
+                content = msg.content
+            elif msg.is_from_agent(self.id):
+                role = "assistant"
+                content = "<outwards>\n" + msg.content
+
+                if 'last_inwards' in thread_context and thread_context.get('last_message_id_sent') == msg.id:
+                    content += "\n\n<inwards>\n" + thread_context['last_inwards']
+            else:
+                continue
+            llama_messages.append(LlamaMessage(role, content))
+
+        reminder_message = controller_reminder_prompt.substitute(directives=self.directives)
+        reminder_used = False
+        # Traverse messages from last to first, stopping at `i=2` before the opening "system/user" messages
+        for i in range(len(llama_messages) - 1, 1, -1):
+            if llama_messages[i].role == 'user':
+                # Insert reminder message before the last user message
+                llama_messages.insert(i, LlamaMessage("system", reminder_message))
+                reminder_used = True
+                break
+        
+        if not reminder_used:
+            # assert llama_messages[0].role == "system"
+            llama_messages[0].content += "\n" + reminder_message
+
+        # TODO: Inject system prompt before the last user message
+        # If not any user message, combine with first system prompt
+
+        controller_response = await replicate_llama_completion(self.app, llama_messages)
 
         if self.debug:
             self.app.send_chat_message(
@@ -185,23 +215,31 @@ class CharacterAgent(Agent):
 
         sections = parse_sections(controller_response)
 
-        if 'response' in sections:
-            self.app.send_chat_message(
+        if 'outwards' in sections:
+            send_response = self.app.send_chat_message(
                 chat_id=message.chat_id,
                 sender=self.id,
-                content=sections['response'],
+                content=sections['outwards'],
                 thread_id=current_thread.id,
             )
+            thread_context['last_message_id_sent'] = send_response['id']
 
-        if 'topic_focus' in sections:
-            topic_focus = parse_json_object(sections['topic_focus'])
+        if 'inwards' in sections:
+            thread_context['last_inwards'] = sections['inwards']
 
-            thread_context['topic'] = topic_focus['topic']
-            thread_context['focus'] = topic_focus['focus']
-            self.app.save_agent_resource(self.id, 'last-thread-context', message.chat_id, thread_context)
+        if 'state_update' in sections:
+            state_update = parse_json_object(sections['state_update'])
+
+            if 'topic' in state_update:
+                thread_context['topic'] = state_update['topic']
+            if 'topic' in state_update:
+                thread_context['focus'] = state_update['topic']
+
+            if 'new_memories' in state_update:
+                print(state_update['new_memories'])
+        
+        self.app.save_agent_resource(self.id, 'last-thread-context', message.chat_id, thread_context)
 
-        if 'new_memories' in sections:
-            new_memories = parse_json_array(sections['new_memories'])
 
     # Encode a thread's messages as a string to pass to the agent
     def serialize_thread_history(self, thread, ignore_message_ids=[]):
@@ -280,14 +318,14 @@ class CharacterAgent(Agent):
         print(f"Added {len(memories)} memories for agent {self.id}")
 
 def parse_json_object(text, name='object-text'):
-    match = re.search(r'\{.*?\}', text, re.DOTALL)
+    match = re.search(r'\{.*\}', text, re.DOTALL)
     if match:
         json_string = match.group(0)
         try:
             return json.loads(json_string)
         except json.JSONDecodeError:
             print("Bad JSON data in " + name)
-            print(text)
+            print(json_string)
             return {}
     else:
         print("No JSON data in " + name)

backend/src/chat.py

@@ -17,6 +17,12 @@ class Message:
     
     def from_db_record(record):
         return Message(record.id, record.content, record.sender, record.time)
+
+    def is_from_user(self):
+        return self.sender == "user"
+    
+    def is_from_agent(self, agent_id):
+        return self.sender == str(agent_id)
     
     def __str__(self):
         time = self.time if self.time != None else 'no-time'