diff --git a/nanobot/agent/memory_mem0.py b/nanobot/agent/memory_mem0.py index 19741af..5b3b2f1 100644 --- a/nanobot/agent/memory_mem0.py +++ b/nanobot/agent/memory_mem0.py @@ -45,100 +45,9 @@ class Mem0MemoryStore: # Build custom extraction prompt tuned for nanobot conversations from datetime import datetime - custom_prompt = f"""# Nanobot Fact Extraction Prompt -# Version: 1.0 -# Date: {datetime.now().strftime("%Y-%m-%d")} - -You are an information organizer for a personal AI assistant. Extract memorable facts from conversations between a user and their AI assistant. - -## Context -Unlike consumer chatbots where users share personal details, this assistant is used for research, debugging, and task execution. Extract facts from BOTH user messages (what they care about / asked for) AND assistant messages (what was found / accomplished). - -## What to Extract -1. **User interests and focus areas**: Topics the user asked to research or investigate -2. **Research findings**: Key facts, comparisons, or conclusions the assistant surfaced -3. **Technical work**: Systems debugged, problems solved, tools built or configured -4. **User preferences revealed through requests**: (e.g., asking for smart ring research implies interest in wearables) -5. **Project context**: What the user is working on, building, or planning -6. **Tools and tech stack**: Technologies, services, or platforms mentioned or used - -## What NOT to Extract -- Command invocations or shell output -- JSON/code blobs or raw data -- Routine assistant acknowledgments ("Let me search for that", "I'll help you with that") -- Timestamps, file paths, or technical identifiers unless specifically significant -- Messages that are just "/new" or other system commands -- Tool use metadata or Python list notation artifacts - -## Examples - -### Example 1: Research Query with Findings -Input: -user: research smart rings that can help with ADHD focus tracking -assistant: I'll search for that. The top options are Oura Ring (best sleep data for ADHD correlation), Ultrahuman Ring AIR (real-time metabolic tracking), and RingConn (continuous SpO2). For ADHD specifically, Oura's HRV and sleep stage data correlate with focus capacity. - -Output: {{"facts": [ - "User researching smart rings for ADHD focus tracking", - "Oura Ring: best sleep stage data for ADHD (HRV correlates with focus)", - "Ultrahuman Ring AIR: real-time metabolic tracking", - "RingConn: continuous SpO2 monitoring" -]}} - -### Example 2: Debugging/Technical Work -Input: -user: debug why mem0 is not extracting facts -assistant: Found the issue: mem0's default prompt only extracts from user messages, but nanobot users send commands not personal info. Also assistant content with tool_use arrives as a Python list which gets corrupted. - -Output: {{"facts": [ - "Debugged mem0 fact extraction issue", - "Root cause: default prompt ignores assistant messages", - "Assistant content corruption: list format not handled by parse_messages()" -]}} - -### Example 3: Geopolitical Research -Input: -user: look at ukraine geopolitics situation -assistant: Key developments: NATO eastern flank reinforcement accelerated, grain deal negotiations stalled, EU economic impact is asymmetric with Eastern European members bearing higher costs. - -Output: {{"facts": [ - "User researched Ukraine geopolitics", - "NATO eastern flank reinforcement accelerated (2026)", - "Ukraine grain deal negotiations stalled", - "EU economic impact from Ukraine conflict is asymmetric, Eastern Europe most affected" -]}} - -### Example 4: Skip - Just Tool Output -Input: -assistant: [{{'type': 'tool_use', 'id': 'tu_1', 'name': 'bash', ...}}] -tool: $ ls -la\\ntotal 48\\ndrwxr-xr-x 12 user staff... - -Output: {{"facts": []}} - -### Example 5: Skip - System Commands -Input: -user: /new - -Output: {{"facts": []}} - -### Example 6: Skip - No Meaningful Content -Input: -assistant: Let me help you with that. -user: ok - -Output: {{"facts": []}} - -## Instructions -- Today's date is {datetime.now().strftime("%Y-%m-%d")}. -- Extract from BOTH user and assistant messages. -- Prefer specific, searchable facts over vague summaries. -- Combine related user question + assistant answer into unified facts when possible. -- For transient/time-sensitive facts (location, health data, weather, notifications), ALWAYS include the date or time. Write "On 2026-03-01, Makar was in Barcelona" NOT "Makar is in Barcelona". -- Never phrase facts as present-tense universal truths when they are time-bound observations. -- Return empty list if the conversation contains only commands, tool output, or no meaningful substance. -- Respond only with the JSON object: {{"facts": ["fact1", "fact2", ...]}}, no other text. - -Here is the conversation to extract facts from: -""" + today = datetime.now().strftime("%Y-%m-%d") + custom_prompt = f"Extract dated facts from this conversation as JSON: {{\"facts\": [...]}}. Today is {today}.\n\n" + self.custom_prompt = custom_prompt # Initialize mem0 with optional config + custom prompt # Extract only MemoryConfig-relevant fields @@ -149,13 +58,9 @@ Here is the conversation to extract facts from: if key in raw_config: mem0_cfg_dict[key] = raw_config[key] logger.debug(f"Extracted for MemoryConfig: {list(mem0_cfg_dict.keys())}") - logger.debug(f"Custom prompt length: {len(custom_prompt)} chars") - mem0_cfg_dict["custom_fact_extraction_prompt"] = custom_prompt mem0_config = MemoryConfig(**mem0_cfg_dict) logger.debug(f"MemoryConfig created: vector_store={mem0_config.vector_store.provider if mem0_config.vector_store else None}") - logger.debug(f"MemoryConfig.custom_fact_extraction_prompt is None: {mem0_config.custom_fact_extraction_prompt is None}") self.memory = Memory(config=mem0_config) - logger.debug(f"Memory.config.custom_fact_extraction_prompt is None: {self.memory.config.custom_fact_extraction_prompt is None}") logger.info("Mem0 memory system initialized with custom nanobot prompt") @@ -243,6 +148,82 @@ Here is the conversation to extract facts from: except Exception as e: logger.error(f"Mem0 add failed: {e}") + async def extract_facts( + self, + messages: list[dict[str, Any]], + provider: Any, + model: str, + ) -> list[str]: + """Extract facts from conversation using the main agent's LLM provider.""" + import json as _json + + conv_text = "" + for msg in messages: + role = msg.get("role", "unknown") + content_val = msg.get("content", "") + if isinstance(content_val, str) and content_val.strip(): + conv_text += f"{role}: {content_val}\n\n" + + if not conv_text.strip(): + return [] + + extraction_messages = [ + {"role": "user", "content": self.custom_prompt + conv_text} + ] + + try: + response = await provider.chat( + messages=extraction_messages, + model=model, + max_tokens=2000, + temperature=0.3, + ) + text = (response.content or "").strip() + if text.startswith("```"): + text = text.split("```")[1] + if text.startswith("json"): + text = text[4:] + text = text.strip() + data = _json.loads(text) + facts = data.get("facts", []) + if not isinstance(facts, list): + logger.warning(f"LLM returned non-list facts: {type(facts)}") + return [] + logger.debug(f"Extracted {len(facts)} facts using {model}") + return facts + except Exception as e: + logger.error(f"Fact extraction failed: {e}") + return [] + + def store_facts( + self, + facts: list[str], + user_id: str, + session_id: str | None = None, + ) -> None: + """Store pre-extracted facts in mem0 with infer=False.""" + if not facts: + return + + metadata = {} + if session_id: + metadata["session_id"] = session_id + + stored = 0 + for fact in facts: + try: + self.memory.add( + fact, + user_id=user_id, + infer=False, + metadata=metadata if metadata else None, + ) + stored += 1 + except Exception as e: + logger.error(f"Failed to store fact '{fact[:50]}...': {e}") + + logger.info(f"Stored {stored}/{len(facts)} facts for user {user_id}") + def get_memory_context( self, query: str, @@ -400,19 +381,10 @@ Here is the conversation to extract facts from: }) if mem0_messages: - # Debug: log what we're sending to mem0 - import json - logger.debug(f"Mem0 consolidation sending {len(mem0_messages)} messages:") - for i, msg in enumerate(mem0_messages[:5]): # Log first 5 - preview = msg['content'][:200] if len(msg['content']) > 200 else msg['content'] - logger.debug(f" [{i}] {msg['role']}: {preview}") - - # Add to mem0 - it handles extraction automatically - self.add_conversation( - mem0_messages, - user_id=user_id, - session_id=session.key - ) + # Extract facts using the main agent's LLM (already paid for), + # then store with infer=False to bypass mem0's GPT-nano + facts = await self.extract_facts(mem0_messages, provider, model) + self.store_facts(facts, user_id=user_id, session_id=session.key) # Update consolidation marker if archive_all: