Memary: Add Memory and Knowledge Graphs to AI Agents

Overview

Memary is a Python-based memory layer for autonomous agents. It stores knowledge in a graph and tracks entities across interactions, giving an agent context about what a user has discussed and how often or recently topics appear. Its goal is to make agent responses more personalized without requiring developers to build these memory mechanisms from scratch.

The repository includes a ReAct-based chat agent and integrations for graph databases, model providers, and tools. Developers can use that agent directly or treat the project as a starting point for adding persistent memory and knowledge-graph retrieval to an agent workflow.

Key Features

• Stores and retrieves entity relationships using a knowledge graph, with FalkorDB and Neo4j listed as database options.
• Tracks a memory stream of encountered entities and an entity store with reference frequency and recency.
• Uses recursive and multi-hop retrieval to assemble relevant graph context for queries.
• Falls back to external search when a query is not found in the graph.
• Summarizes earlier chat history to reduce context-window pressure.
• Supports local Ollama models as well as specified OpenAI models for language and vision tasks.
• Offers default tools for search, vision, location, and stock queries, with support for adding or removing custom tools.
• Supports separate agent contexts through user IDs and, with FalkorDB, multiple graphs.

Use Cases

• Build a personal assistant that adapts responses based on a user's interests and prior conversations.
• Add graph-backed retrieval to an agent that needs to connect entities across multiple facts or interactions.
• Prototype multi-user agents with distinct memory and knowledge contexts.
• Extend the included ReAct agent with custom tools while keeping memory and retrieval in the same application.

Project Facts

• Language: Jupyter Notebook
• License: MIT
• Stars: 2.7k
• Forks: 206
• Topics: agents, knowledge-graph, memory, multiagent-systems, rag, self-improvement
• Archived: No

Getting Started

Install from PyPI with pip install memary. The README specifies Python version <= 3.11.9. Running the included Streamlit app also requires configuring relevant model, database, and tool API credentials. See the README and documentation for setup details.

Alternatives

• Cognithor: Cognithor is a local-first autonomous agent system with six-tier cognitive memory, while Memary provides a memory layer for developers building agents.

Considerations

• The README specifies Python version <= 3.11.9, so newer Python versions may not be supported.
• A working setup may need credentials for model providers, graph databases, and optional tools such as Google Maps or external search.
• The included application uses a ReAct agent, and the README describes it as a demo-oriented implementation intended to be replaced in future versions. Treat it as an integration starting point rather than assuming provider-independent agent support is already available.
• Memory is stored through graph databases or local JSON files, so deployment choices affect persistence and multi-agent setup.