Ask most people what's holding AI back and they will say intelligence. Charles Packer says they have the wrong answer. The gap, he argues, is memory. A language model can pass the bar exam and still forget your name between two sentences. It does not learn from Monday's mistake before repeating it on Tuesday. Packer's entire career is a bet on fixing that one thing.
He is the co-founder and CEO of Letta, the San Francisco company building the memory layer for AI agents - what he calls the operating system that sits above the base models. Before Letta there was MemGPT, the 2023 research project that reframed a large language model as something closer to a computer: a small, fast working memory up front, a deep store behind it, and software that decides what to page in and out. The idea was elegant, a little contrarian, and it spread fast. The open-source repo passed 10,000 stars before the company that would commercialize it had a name.
The fundamental difference between humans and LLMs right now is not intelligence. It is memory. Humans can learn. LLMs cannot.
Letta emerged from stealth in September 2024 with backing that reads like an AI hall of fame: Jeff Dean of Google DeepMind, Clem Delangue of Hugging Face, Cristobal Valenzuela of Runway, and Robert Nishihara of Anyscale all wrote checks. The lead investor, Felicis Ventures, put the round together around a single thesis - that the durable value in AI is not any one model, but the stateful layer that lives above all of them.
Packer spent close to five years as a PhD student at UC Berkeley, working out of two of the most consequential labs in machine learning - the Berkeley AI Research lab (BAIR) and the Sky Computing Lab, the group formerly known as RISELab and AMPLab. His early research wandered through reinforcement learning and autonomous driving before large language models pulled the whole field toward agents.
His Berkeley thesis carried a telling title: building agentic systems in an era of large language models. The framing matters. Packer never treated the model as the product. To him, the LLM is one amazing component inside a larger machine that also needs to plan, act, and remember. MemGPT was the missing memory component, written down as research and then handed to the world.
When the project went viral, the choice in front of him was the kind most researchers only daydream about. He had offers from the biggest names in the industry. He took none of them. Together with co-founder Sarah Wooders, he turned MemGPT into Letta and set up in Jackson Square, San Francisco.
The next frontier in AI is in the stateful layer above the base models - the memory layer, or LLM OS.
Packer's term for agents that actually keep state - they remember across sessions and improve after deployment, not just during training. His blunt take: most things called "agents" today are stateless workflows wearing a nice interface.
A background process that consolidates an agent's memory while it is idle. The agent keeps thinking about your problems between conversations - the digital equivalent of sleeping on it.
An open standard for packaging an agent's memory and configuration into a portable file - a kind of exportable "soul" you can move between systems. The format is deliberately open, in keeping with Letta's stance.
"Would you rather have an AI coworker that makes the same mistake every single week forever, or a human who makes five times the mistakes upfront but never repeats one?"
"Most of what people call agents today are stateless workflows with a pretty interface. They do not have memory. They do not learn. They do not get better."
"Every company is going to have a living digital copy of every customer. That copy will live inside a stateful agent. The question is: what platform does that agent run on?"
"Large language models are just one (amazing) piece of a complete agentic system."
Packer on the thinking behind MemGPT and the road to Letta - why the memory layer, why open, and why now.
Packer wants Letta to be the default platform that every stateful agent runs on - the open memory layer that lets AI remember, learn, and get better long after it ships. If he is right that memory is the real bottleneck, the company sitting on that layer is not building a feature. It is building the ground floor.