Memary: an open-source project to enhance Agent long-term memory using knowledge graphs

General Introduction

Memary is an innovative open source project focused on providing long-term memory management solutions for autonomous intelligences. The project helps intelligences break through the limitations of traditional context windows to achieve a smarter interaction experience through knowledge graphs and specialized memory modules.Memary uses an automated memory generation mechanism that can automatically update memories during intelligence interactions and display these memories through a unified dashboard. The system supports multiple model configurations, including locally running Llama and LLaVA models, as well as GPT models in the cloud. In addition, Memary provides multi-graph support, allowing developers to create independent instances of intelligences for different users, enabling personalized memory management.

Memary Overall Architecture

Memary Agent

Memary Knowledge Graphs

Memory Modules

Function List

• Automated Memory Generation and Updating
• Knowledge graph storage and retrieval
• Memory Stream Tracking Entities and Timestamps
• Frequency and timeliness of Entity Knowledge Store (EKS) management
• Recursive Retrieval Methods to Optimize Knowledge Graph Search
• Multi-hop reasoning supports complex query processing
• Customized Tool Extension Support
• Multi-Intelligence Body Mapping Management
• Memory compression and context window optimization
• Topic Extraction and Entity Categorization
• Time line analysis function

Using Help

1. Installation configuration

1.1 Basic requirements

• Python Version Requirements: <= 3.11.9
• Recommended to use a virtual environment for installation

1.2 Installation method
a) Use pip to install.

pip install memary

b) Local installation.

• Create and activate a virtual environment
• Installation dependencies. pip install -r requirements.txt

1.3 Model configuration
Memary supports two modes of operation:

• Local mode (default): Use the Ollama operational model
  • LLM: Llama 3 8B/40B (recommended)
  • Visual model: LLaVA (recommended)
• Cloud mode:
  • LLM: gpt-3.5-turbo
  • Vision model: gpt-4-vision-preview

2. Environmental preparation

2.1 Configuring the .env File

OPENAI_API_KEY="YOUR_API_KEY"
PERPLEXITY_API_KEY="YOUR_API_KEY"
GOOGLEMAPS_API_KEY="YOUR_API_KEY"
ALPHA_VANTAGE_API_KEY="YOUR_API_KEY"
# Database Configuration (either one or the other):
FALKORDB_URL="falkor://[[username]:[password]]@[falkor_host_url]:port"
or
NEO4J_PW="YOUR_NEO4J_PW"
NEO4J_URL="YOUR_NEO4J_URL"

2.2 Update user configuration

• compiler streamlit_app/data/user_persona.txt Setting user characteristics
• Optional: Modify streamlit_app/data/system_persona.txt Adjustment of system characteristics

3. Basic use

3.1 Launching the application

cd streamlit_app
streamlit run app.py

3.2 Code Examples

from memary.agent.chat_agent import ChatAgent
# Initialize the ChatAgent
chat_agent = ChatAgent(
"Personal Agent",
memory_stream_json,
entity_knowledge_store_json,
system_persona_txt,
system_persona_txt, user_persona_txt,
past_chat_json.
)
# Adding a custom tool
def multiply(a: int, b: int) -> int.
"""Multiply Calculation Tool""""
return a * b
chat_agent.add_tool({"multiply": multiply})
# Remove the tool
chat_agent.remove_tool("multiply")

4. Multi-intelligence configuration

Applies when using the FalkorDB database:

# Personal Intelligence of User A
chat_agent_user_a = ChatAgent(
"Personal Agent",
memory_stream_json_user_a,
entity_knowledge_store_json_user_a,
system_persona_txt_user_a,
user_persona_txt_user_a,
past_chat_json_user_a,
user_id='user_a_id'
)
# Personal Intelligence for User B
chat_agent_user_b = ChatAgent(
"Personal Agent",
memory_stream_json_user_b,
entity_knowledge_store_json_user_b,
system_persona_txt_user_b,
user_persona_txt_user_b,
past_chat_json_user_b,
user_id='user_b_id'
)

5. Memory management features

5.1 Memory Stream

• Automatically capture all entities and their timestamps
• Support for timeline analysis
• Theme Extraction Function

5.2 Entity Knowledge Store

• Tracking entity citation frequency and timeliness
• Entity Relevance Ranking
• entity classification function
• Time-varying analysis

5.3 Knowledge Mapping Functions

• Recursive Search Optimization
• Multi-hop reasoning support
• Automatic update mechanism