MemoryOS is designed to provide a memory operating system for personalized AI agents, enabling more coherent, personalized, and context-aware interactions. Drawing inspiration from memory management principles in operating systems, it adopts a hierarchical storage architecture with four core modules: Storage, Updating, Retrieval, and Generation, to achieve comprehensive and efficient memory management. On the LoCoMo benchmark, the model achieved average improvements of 49.11% and 46.18% in F1 and BLEU-1 scores.
- [2025-05-30]: Initial version of MemoryOS launched! Featuring short-term, mid-term, and long-term persona Memory with automated user profile and knowledge updating.
memoryos/
├── __init__.py # Initializes the MemoryOS package
├── __pycache__/ # Python cache directory (auto-generated)
├── long_term.py # Manages long-term persona memory (user profile, knowledge)
├── memoryos.py # Main class for MemoryOS, orchestrating all components
├── mid_term.py # Manages mid-term memory, consolidating short-term interactions
├── prompts.py # Contains prompts used for LLM interactions (e.g., summarization, analysis)
├── retriever.py # Retrieves relevant information from all memory layers
├── short_term.py # Manages short-term memory for recent interactions
├── updater.py # Processes memory updates, including promoting information between layers
└── utils.py # Utility functions used across the library
- Initialization:
Memoryosis initialized with user and assistant IDs, API keys, data storage paths, and various capacity/threshold settings. It sets up dedicated storage for each user and assistant. - Adding Memories: User inputs and agent responses are added as QA pairs. These are initially stored in short-term memory.
- Short-Term to Mid-Term Processing: When short-term memory is full, the
Updatermodule processes these interactions, consolidating them into meaningful segments and storing them in mid-term memory. - Mid-Term Analysis & LPM Updates: Mid-term memory segments accumulate "heat" based on factors like visit frequency and interaction length. When a segment's heat exceeds a threshold, its content is analyzed:
- User profile insights are extracted and used to update the long-term user profile.
- Specific user facts are added to the user's long-term knowledge.
- Relevant information for the assistant is added to the assistant's long-term knowledge base.
- Response Generation: When a user query is received:
- The
Retrievermodule fetches relevant context from short-term history, mid-term memory segments, the user's profile & knowledge, and the assistant's knowledge base. - This comprehensive context is then used, along with the user's query, to generate a coherent and informed response via an LLM.
- The
- Python >= 3.10
- pip install -i https://pypi.org/simple/ MemoryOS-BaiJia
conda create -n MemoryOS python=3.10
conda activate MemoryOS
pip install -i https://pypi.org/simple/ MemoryOS-BaiJiaimportosfrommemoryosimportMemoryos# --- Basic Configuration ---USER_ID="demo_user"ASSISTANT_ID="demo_assistant"API_KEY="YOUR_OPENAI_API_KEY"# Replace with your keyBASE_URL=""# Optional: if using a custom OpenAI endpointDATA_STORAGE_PATH="./simple_demo_data"LLM_MODEL="gpt-4o-mini"defsimple_demo():
print("MemoryOS Simple Demo")
# 1. Initialize MemoryOSprint("Initializing MemoryOS...")
try:
memo=Memoryos(
user_id=USER_ID,
openai_api_key=API_KEY,
openai_base_url=BASE_URL,
data_storage_path=DATA_STORAGE_PATH,
llm_model=LLM_MODEL,
assistant_id=ASSISTANT_ID,
short_term_capacity=7, mid_term_heat_threshold=5, retrieval_queue_capacity=7,
long_term_knowledge_capacity=100
)
print("MemoryOS initialized successfully!\n")
exceptExceptionase:
print(f"Error: {e}")
return# 2. Add some basic memoriesprint("Adding some memories...")
memo.add_memory(
user_input="Hi! I'm Tom, I work as a data scientist in San Francisco.",
agent_response="Hello Tom! Nice to meet you. Data science is such an exciting field. What kind of data do you work with?"
)
test_query="What do you remember about my job?"print(f"User: {test_query}")
response=memo.get_response(
query=test_query,
)
print(f"Assistant: {response}")
if__name__=="__main__":
simple_demo()Contributions are welcome! Please feel free to submit issues or pull requests.
If you want to read more details, please click here: Read the Full Paper
If you find this project useful, please consider citing our paper:
The 'Memory OS of AI Agent' paper will be available on arXiv soon.百家AI是北京邮电大学白婷副教授指导的研究小组, 致力于为硅基人类打造情感饱满、记忆超凡的大脑。
合作与建议：baiting@bupt.edu.cn
欢迎关注百家Agent公众号