Memary provides an open-source memory layer for autonomous agents, aiming to emulate human memory to enhance agent reasoning and task execution. It's designed for developers building advanced AI agents, offering a structured way to manage and leverage agent interactions and knowledge.

How It Works

Memary integrates with existing agents by automatically updating memory as the agent interacts. It utilizes a graph database (FalkorDB or Neo4j) for knowledge storage, with LlamaIndex used for adding nodes. For external queries, it leverages Perplexity (mistral-7b-instruct) or OpenAI. The system employs recursive retrieval and multi-hop reasoning for efficient knowledge graph querying, reducing latency by focusing on relevant subgraphs. Memory is managed through a Memory Stream (tracking entities and timestamps) and an Entity Knowledge Store (tracking entity frequency and recency).

Quick Start & Requirements

• Install: pip install memary
• Python Version: <= 3.11.9
• Local Models: Requires Ollama for LLMs (e.g., Llama 3 8B/40B) and Vision models (e.g., LLaVA).
• API Keys: OpenAI, Perplexity, Google Maps, Alpha Vantage, FalkorDB, or Neo4j.
• Setup: Follow Ollama instructions for local models. Configure .env with API keys and database credentials.
• Demo: Run streamlit run app.py from the streamlit_app directory.
• Docs: OpenAI key, FalkorDB, Neo4j, Perplexity key, Google Maps, Alpha Vantage.

Highlighted Details

• Supports multi-graph capabilities with FalkorDB for managing multiple agents.
• Allows adding and removing custom tools for agent functionality.
• Memory module design influenced by Microsoft Research's K-LaMP.
• Features recursive retrieval and multi-hop reasoning for efficient KG querying.

Maintenance & Community

No specific contributors, sponsorships, or community links (Discord/Slack) are mentioned in the README.

Licensing & Compatibility

• License: MIT License.
• Compatibility: Permissive for commercial use and closed-source linking.

Limitations & Caveats

The project is noted to use a ReAct agent for demonstration purposes, with plans to remove it in future releases to support any agent type. The README specifies Python version <= 3.11.9, which might be a constraint for users on newer Python versions.