Method query

methods/memrl_agent.py:725–871

Query using MemRL's value-driven retrieval. This method: 1. Uses MemRL's retrieve_query for value-aware memory retrieval 2. Applies Q-value based ranking with ε-greedy exploration 3. Constructs prompt with retrieved memories 4. Generates response using the tr

(
        self,
        question: str,
        system_message: Optional[str] = None,
        **kwargs
    )

        )

    def query(
        self,
        question: str,
        system_message: Optional[str] = None,
        **kwargs
    ) -> AgentResponse:
        """Query using MemRL's value-driven retrieval.

        This method:
        1. Uses MemRL's retrieve_query for value-aware memory retrieval
        2. Applies Q-value based ranking with ε-greedy exploration
        3. Constructs prompt with retrieved memories
        4. Generates response using the tracked LLM client
        """
        start_time = time.time()

        # Truncate question if needed
        bounded_question = self._truncate_to_tokens(question, self.max_question_tokens)

        # Retrieve memories using MemRL's value-driven retrieval
        retrieved_memories: List[Dict[str, Any]] = []
        retrieval_result: Dict[str, Any] = {}

        try:
            # Use MemRL's retrieve_query which combines similarity and Q-value
            result = self._memory_service.retrieve_query(
                task_description=bounded_question,
                k=self.candidate_top_k,
                threshold=self.similarity_threshold,
            )

            # retrieve_query returns (result_dict, sim_list) tuple
            retrieval_result, _ = result

            # Extract selected memories from result
            selected = retrieval_result.get("selected", [])
            for mem in selected[:self.retrieve_num]:
                # Extract content from memory object
                content = self._extract_memory_content(mem)

                retrieved_memories.append({
                    "memory": content[:self.query_memory_item_tokens * 4],  # Rough char limit
                    "memory_id": mem.get("memory_id"),
                    "similarity": mem.get("similarity", 0.0),
                    "q_value": mem.get("q_estimate", self.initial_q),
                    "score": mem.get("score", 0.0),
                    "type": "memrl_value_driven",
                })

        except Exception as e:
            logger.warning(f"[MemRL] Retrieval failed: {e}, falling back to empty context")
            retrieval_result = {"selected": [], "candidates": [], "simmax": 0.0}

        retrieval_time = time.time() - start_time

        # Calculate token budget for memory context
        system_tokens = self._llm_client.count_tokens(system_message) if system_message else 0
        reserved_tokens = self.max_tokens + 500