Felicis
- MemGPT helps users "build LLM agents with long term memory and custom tools."
- It is an agentic framework focused on increasing the memory of large language models.
- The MemGPT project has over 11,000 stars on GitHub.
MemGPT, a toolkit for developing AI agents, has raised a $10 million round at a $70 million valuation, Business Insider has learned.
Felicis Ventures led the round, four sources familiar said. This is the company's first round, a source said.
MemGPT helps users "build LLM agents with long term memory and custom tools," according to its associated GitHub page. The MemGPT project has over 11,000 stars on GitHub, indicating the project's popularity among developers.
MemGPT and Felicis Ventures did not respond to requests for comment.
MemGPT is an agent tuned for short-term and long-term memory, a source said. While short-term memory gives LLM agents the immediate context for a single session, long-term memory serves as a knowledge base from which agents can recall prior context.
The startup's website highlights MemGPT's key features, including adaptive and long-term memory, extensive context windows, and unlimited data. Developers can also use MemGPT to build multiple agents for multiple users.
Sarah Wooders and Charles Parker, both UC Berkeley Ph.D. graduates, founded the company and collaborated on the research paper behind it. According to the paper, large language models' capabilities are limited by their context windows — the amount of information an LLM can process at once. For instance, LLMs have trouble analyzing hundreds of large documents.
MemGPT developed an agentic framework, or a toolkit to build AI agents, to address this issue, focused on memory management. This framework enables AI agents to process more information than they typically could. It uses an approach inspired by "traditional operating systems," enabling LLMs to analyze large documents and remember long conversations, even when the information exceeds the model's usual processing limits.
Emergence, which raised about $100 million, is a multi-agent framework for automating enterprise knowledge work. Other startups, such as Crew AI and Phidata, also aim to accelerate multi-agent systems, which are increasingly a hot investment area.
Big Tech has also dabbled in the space. Microsoft's AutoGen and Semantic Kernel are open-source toolkits for AI agent development. Amazon Bedrock enables developers to access and build on top of various foundation models through a single API.