The Memory Layer for AI Agents - Drop-in memory infrastructure for AI agents and apps. Context that persists. Built for production.
Python agentic-memoryagentic-memory-systemagentsaiai-agentschatgptgenaillmlong-term-memorymemorymemory-managementpythonragstate-management
Add: Json Parsing to solve Hallucination Errors (#3013)
Akshat Jain2bb0653e679c
diff --git a/evaluation/metrics/llm_judge.py b/evaluation/metrics/llm_judge.pyindex 4d0ec376a..8dbae80b0 100644--- a/evaluation/metrics/llm_judge.py+++ b/evaluation/metrics/llm_judge.py@@ -4,6 +4,7 @@ from collections import defaultdictimport numpy as npfrom openai import OpenAI+from mem0.memory.utils import extract_jsonclient = OpenAI()@@ -22,7 +23,7 @@ The generated answer might be much longer, but you should be generous with yourFor time related questions, the gold answer will be a specific date, month, year, etc. The generated answer might be much longer or use relative time references (like "last Tuesday" or "next month"), but you should be generous with your grading - as long as it refers to the same date or time period as the gold answer, it should be counted as CORRECT. Even if the format differs (e.g., "May 7th" vs "7 May"), consider it CORRECT if it's the same date.-Now it’s time for the real question:+Now it's time for the real question:Question: {question}Gold answer: {gold_answer}Generated answer: {generated_answer}@@ -49,7 +50,7 @@ def evaluate_llm_judge(question, gold_answer, generated_answer):response_format={"type": "json_object"},temperature=0.0,)- label = json.loads(response.choices[0].message.content)["label"]+ label = json.loads(extract_json(response.choices[0].message.content))["label"]return 1 if label == "CORRECT" else 0diff --git a/mem0/embeddings/aws_bedrock.py b/mem0/embeddings/aws_bedrock.pyindex 807764c1f..1679aeec0 100644--- a/mem0/embeddings/aws_bedrock.py+++ b/mem0/embeddings/aws_bedrock.py@@ -11,6 +11,7 @@ import numpy as npfrom mem0.configs.embeddings.base import BaseEmbedderConfigfrom mem0.embeddings.base import EmbeddingBase+from mem0.memory.utils import extract_jsonclass AWSBedrockEmbedding(EmbeddingBase):@@ -74,7 +75,7 @@ class AWSBedrockEmbedding(EmbeddingBase):contentType="application/json",)- response_body = json.loads(response.get("body").read())+ response_body = json.loads(extract_json(response.get("body").read()))if provider == "cohere":embeddings = response_body.get("embeddings")[0]diff --git a/mem0/llms/aws_bedrock.py b/mem0/llms/aws_bedrock.pyindex dde66f30f..91a2806f5 100644--- a/mem0/llms/aws_bedrock.py+++ b/mem0/llms/aws_bedrock.py@@ -10,6 +10,7 @@ except ImportError:from mem0.configs.llms.base import BaseLlmConfigfrom mem0.llms.base import LLMBase+from mem0.memory.utils import extract_jsonPROVIDERS = ["ai21", "amazon", "anthropic", "cohere", "meta", "mistral", "stability", "writer"]@@ -101,7 +102,7 @@ class AWSBedrockLLM(LLMBase):return processed_responseresponse_body = response.get("body").read().decode()- response_json = json.loads(response_body)+ response_json = json.loads(extract_json(response_body))return response_json.get("content", [{"text": ""}])[0].get("text", "")def _prepare_input(diff --git a/mem0/llms/azure_openai.py b/mem0/llms/azure_openai.pyindex a7f1fdaf1..1dcb5f3fb 100644--- a/mem0/llms/azure_openai.py+++ b/mem0/llms/azure_openai.py@@ -6,6 +6,7 @@ from openai import AzureOpenAIfrom mem0.configs.llms.base import BaseLlmConfigfrom mem0.llms.base import LLMBase+from mem0.memory.utils import extract_jsonclass AzureOpenAILLM(LLMBase):@@ -53,7 +54,7 @@ class AzureOpenAILLM(LLMBase):processed_response["tool_calls"].append({"name": tool_call.function.name,- "arguments": json.loads(tool_call.function.arguments),+ "arguments": json.loads(extract_json(tool_call.function.arguments)),})diff --git a/mem0/llms/deepseek.py b/mem0/llms/deepseek.pyindex 46a805f0c..85a0417a6 100644--- a/mem0/llms/deepseek.py+++ b/mem0/llms/deepseek.py@@ -6,6 +6,7 @@ from openai import OpenAIfrom mem0.configs.llms.base import BaseLlmConfigfrom mem0.llms.base import LLMBase+from mem0.memory.utils import extract_jsonclass DeepSeekLLM(LLMBase):@@ -41,7 +42,7 @@ class DeepSeekLLM(LLMBase):processed_response["tool_calls"].append({"name": tool_call.function.name,- "arguments": json.loads(tool_call.function.arguments),+ "arguments": json.loads(extract_json(tool_call.function.arguments)),})diff --git a/mem0/llms/groq.py b/mem0/llms/groq.pyindex e970a1ee0..cc8733d5c 100644--- a/mem0/llms/groq.py+++ b/mem0/llms/groq.py@@ -9,6 +9,7 @@ except ImportError:from mem0.configs.llms.base import BaseLlmConfigfrom mem0.llms.base import LLMBase+from mem0.memory.utils import extract_jsonclass GroqLLM(LLMBase):@@ -43,7 +44,7 @@ class GroqLLM(LLMBase):processed_response["tool_calls"].append({"name": tool_call.function.name,- "arguments": json.loads(tool_call.function.arguments),+ "arguments": json.loads(extract_json(tool_call.function.arguments)),})diff --git a/mem0/llms/litellm.py b/mem0/llms/litellm.pyindex d5896ff80..3a5ef60c6 100644--- a/mem0/llms/litellm.py+++ b/mem0/llms/litellm.py@@ -8,6 +8,7 @@ except ImportError:from mem0.configs.llms.base import BaseLlmConfigfrom mem0.llms.base import LLMBase+from mem0.memory.utils import extract_jsonclass LiteLLM(LLMBase):@@ -39,7 +40,7 @@ class LiteLLM(LLMBase):processed_response["tool_calls"].append({"name": tool_call.function.name,- "arguments": json.loads(tool_call.function.arguments),+ "arguments": json.loads(extract_json(tool_call.function.arguments)),})diff --git a/mem0/llms/openai.py b/mem0/llms/openai.pyindex 7fc3ff4db..8f3cf807a 100644--- a/mem0/llms/openai.py+++ b/mem0/llms/openai.py@@ -7,6 +7,7 @@ from openai import OpenAIfrom mem0.configs.llms.base import BaseLlmConfigfrom mem0.llms.base import LLMBase+from mem0.memory.utils import extract_jsonclass OpenAILLM(LLMBase):@@ -62,7 +63,7 @@ class OpenAILLM(LLMBase):processed_response["tool_calls"].append({"name": tool_call.function.name,- "arguments": json.loads(tool_call.function.arguments),+ "arguments": json.loads(extract_json(tool_call.function.arguments)),})diff --git a/mem0/llms/together.py b/mem0/llms/together.pyindex 922a30d22..d2af10c18 100644--- a/mem0/llms/together.py+++ b/mem0/llms/together.py@@ -9,6 +9,7 @@ except ImportError:from mem0.configs.llms.base import BaseLlmConfigfrom mem0.llms.base import LLMBase+from mem0.memory.utils import extract_jsonclass TogetherLLM(LLMBase):@@ -43,7 +44,7 @@ class TogetherLLM(LLMBase):processed_response["tool_calls"].append({"name": tool_call.function.name,- "arguments": json.loads(tool_call.function.arguments),+ "arguments": json.loads(extract_json(tool_call.function.arguments)),})diff --git a/mem0/llms/utils/__init__.py b/mem0/llms/utils/__init__.pydeleted file mode 100644index e69de29bb..000000000diff --git a/mem0/llms/utils/functions.py b/mem0/llms/utils/functions.pydeleted file mode 100644index e69de29bb..000000000diff --git a/mem0/llms/vllm.py b/mem0/llms/vllm.pyindex e522068ba..1f0fc8227 100644--- a/mem0/llms/vllm.py+++ b/mem0/llms/vllm.py@@ -4,6 +4,7 @@ from typing import Dict, List, Optionalfrom mem0.configs.llms.base import BaseLlmConfigfrom mem0.llms.base import LLMBase+from mem0.memory.utils import extract_jsonclass VllmLLM(LLMBase):@@ -39,7 +40,7 @@ class VllmLLM(LLMBase):for tool_call in response.choices[0].message.tool_calls:processed_response["tool_calls"].append({"name": tool_call.function.name,- "arguments": json.loads(tool_call.function.arguments),+ "arguments": json.loads(extract_json(tool_call.function.arguments)),})return processed_responsediff --git a/mem0/memory/utils.py b/mem0/memory/utils.pyindex b083508df..9018d5461 100644--- a/mem0/memory/utils.py+++ b/mem0/memory/utils.py@@ -46,6 +46,20 @@ def remove_code_blocks(content: str) -> str:return match.group(1).strip() if match else content.strip()+def extract_json(text):+ """+ Extracts JSON content from a string, removing enclosing triple backticks and optional 'json' tag if present.+ If no code block is found, returns the text as-is.+ """+ text = text.strip()+ match = re.search(r"```(?:json)?\s*(.*?)\s*```", text, re.DOTALL)+ if match:+ json_str = match.group(1)+ else:+ json_str = text # assume it's raw JSON+ return json_str++def get_image_description(image_obj, llm, vision_details):"""Get the description of the imagediff --git a/mem0/vector_stores/azure_ai_search.py b/mem0/vector_stores/azure_ai_search.pyindex 6acd2e40d..7e06cff71 100644--- a/mem0/vector_stores/azure_ai_search.py+++ b/mem0/vector_stores/azure_ai_search.py@@ -6,6 +6,7 @@ from typing import List, Optionalfrom pydantic import BaseModelfrom mem0.vector_stores.base import VectorStoreBase+from mem0.memory.utils import extract_jsontry:from azure.core.credentials import AzureKeyCredential@@ -233,7 +234,7 @@ class AzureAISearch(VectorStoreBase):results = []for result in search_results:- payload = json.loads(result["payload"])+ payload = json.loads(extract_json(result["payload"]))results.append(OutputData(id=result["id"], score=result["@search.score"], payload=payload))return results@@ -288,7 +289,8 @@ class AzureAISearch(VectorStoreBase):result = self.search_client.get_document(key=vector_id)except ResourceNotFoundError:return None- return OutputData(id=result["id"], score=None, payload=json.loads(result["payload"]))+ payload = json.loads(extract_json(result["payload"]))+ return OutputData(id=result["id"], score=None, payload=payload)def list_cols(self) -> List[str]:"""@@ -335,7 +337,7 @@ class AzureAISearch(VectorStoreBase):search_results = self.search_client.search(search_text="*", filter=filter_expression, top=limit)results = []for result in search_results:- payload = json.loads(result["payload"])+ payload = json.loads(extract_json(result["payload"]))results.append(OutputData(id=result["id"], score=result["@search.score"], payload=payload))return [results]diff --git a/mem0/vector_stores/redis.py b/mem0/vector_stores/redis.pyindex d28199753..8e85055f1 100644--- a/mem0/vector_stores/redis.py+++ b/mem0/vector_stores/redis.py@@ -12,6 +12,7 @@ from redisvl.query import VectorQueryfrom redisvl.query.filter import Tagfrom mem0.vector_stores.base import VectorStoreBase+from mem0.memory.utils import extract_jsonlogger = logging.getLogger(__name__)@@ -175,7 +176,7 @@ class RedisDB(VectorStoreBase):else {}),**{field: result[field] for field in ["agent_id", "run_id", "user_id"] if field in result},- **{k: v for k, v in json.loads(result["metadata"]).items()},+ **{k: v for k, v in json.loads(extract_json(result["metadata"])).items()},},)for result in results@@ -219,7 +220,7 @@ class RedisDB(VectorStoreBase):else {}),**{field: result[field] for field in ["agent_id", "run_id", "user_id"] if field in result},- **{k: v for k, v in json.loads(result["metadata"]).items()},+ **{k: v for k, v in json.loads(extract_json(result["metadata"])).items()},}return MemoryResult(id=result["memory_id"], payload=payload)@@ -286,7 +287,7 @@ class RedisDB(VectorStoreBase):for field in ["agent_id", "run_id", "user_id"]if field in result.__dict__},- **{k: v for k, v in json.loads(result["metadata"]).items()},+ **{k: v for k, v in json.loads(extract_json(result["metadata"])).items()},},)for result in results.docs