Here’s a question worth sitting with for a second.
Would you rather work with someone who makes the exact same mistake every Monday morning for the next three years, or someone who stumbles five times in month one, then never again?
Most people answer instantly. Charles Packer asks it deliberately.
Thanks for reading! This post is public so feel free to share it.
Because that question is, in his words, the entire reason Letta exists.
Packer is 32. He spent nearly five years deep inside UC Berkeley’s AI research labs. He watched models get smarter, faster, cheaper. He watched demos get more impressive. He watched the hype cycle churn.
And he kept noticing the same thing nobody wanted to say out loud.
Every morning, the AI woke up knowing nothing.
Not because the models weren’t good. But because they had no memory. No ability to carry what happened yesterday into today.
No way to get better at working with you specifically, over time. Every conversation from scratch. Every mistake, available to be repeated indefinitely.
“The fundamental difference between humans and LLMs right now is not intelligence,” Packer told a reporter recently. “It is memory. Humans can learn. LLMs cannot.”
He turned down post-PhD offers from the kinds of places that make your LinkedIn look very impressive. Then he went and built Letta instead.
In 2023, Packer and his co-founder Sarah Wooders, now Letta’s CTO, published a paper out of Berkeley’s BAIR lab.
It was called MemGPT. The idea was straightforward: what if you could give a language model something resembling actual persistent memory?
The paper hit. Then kept hitting. The GitHub repo crossed 10,000 stars. Developers started building on it without being asked to.
That’s usually when you know something is real.
By August 2024, Packer and Wooders had raised a $10 million seed round led by Felicis Ventures. The angels who wrote checks included Jeff Dean (Google DeepMind’s chief scientist), Clem Delangue (CEO of HuggingFace), and Cristóbal Valenzuela of Runway. These are not people who invest in demos. They invest in theses they believe.
The thesis here: the next AI battleground isn’t the model. It’s what sits above it. The memory layer. The persistent state. What Packer calls the “LLM OS.”
You’ll hear the word “agent” everywhere in tech right now. Packer doesn’t have much patience for how loosely it gets used.
“Most of what people call agents today are stateless workflows with a pretty interface,” he’s said. “They do not have memory. They do not learn. They do not get better.”
A stateful AI is different. It remembers what happened. It updates its understanding. It improves through experience, the same way a good employee gets sharper after six months on the job.
Letta’s most visible deployment is with Bilt, a loyalty platform. Every Bilt user has dozens of agents tracking their spending behavior, updating in real time, learning what they actually care about.
Not a static recommendation engine. A system that genuinely knows you better in month three than it did in month one.
“Before Letta, Bilt was using static rankings. The same offer whether it was a user’s first week or their third year,” Packer explained. “With stateful agents, the system learns what a person cares about over time. That is not a product improvement. That is a category shift.”
One of Letta’s most interesting technical bets is something they call Sleep-time Compute.
The idea is borrowed, loosely, from how human memory actually works. We don’t process everything in real time. A lot of consolidation happens while we sleep, the brain quietly sorting, organizing, filing experience away.
Letta’s agents do something similar. Between sessions, in the background, they process and consolidate what they’ve learned.
It’s a clean solution to one of the ugliest problems in production AI: context windows fill up, and most systems just start forgetting things.
Early benchmarks on Terminal-Bench 2.0 show a 21% relative performance improvement from agent-learned skills, alongside a 15.7% drop in costs. Those are not small numbers.
Here’s where Packer gets a little sharper.
Letta is open source. Deliberately, principally so, in direct opposition to where most of the major AI providers are heading.
“Developers deserve to know what is actually going into their models,” he’s said. “The moment you let a closed system own your agent’s memory, you have handed over your most valuable asset.”
Letta Code, their memory-first coding agent, is built to compete with tools like Claude Code. The key difference: it learns.
A developer who has worked with the same Letta Code agent for three weeks has an agent that knows their codebase, remembers past corrections, and has internalized how they like to work.
Packer has been running his own Letta Code agent for over a month, the same one he used to actually build Letta Code itself. He says the compounding is real. Every correction sticks.
They also shipped something called Agent File (.af), an open standard that packages everything making a stateful agent useful into a single portable file. Memory, tool configurations, conversation history, model settings. Think Docker, but for an AI’s brain. Back it up, version it, move it to a different model provider without starting over.
Packer’s long-term bet is not complicated. He thinks every company will eventually have a living digital copy of every customer, running inside a stateful agent, getting smarter with every interaction.
The question, the only question that really matters for Letta, is whose infrastructure that runs on.
He’s not threatened by the big labs shipping new features. When Anthropic released new programmatic tool-calling capabilities, Letta published a full breakdown within days and added support across all model providers within the week.
That’s the posture: stay model-agnostic, move fast, and let the value live in the memory layer no matter who wins the model race.
“We are not betting on any one lab winning,” he says. “We are betting on memory being important regardless of who wins.”
That, honestly, feels like a safe bet.
Kelvorn is now accepting feature applications.
If you’re a founder, builder, creator, or operator with a story worth telling, we want to hear from you. Quiet builders, fast scalers, people doing meaningful work that rarely gets the attention it deserves. All of it counts.
If this one stayed with you, send it to someone who’s deep in the build right now. They’ll appreciate it more than you think.
And if you found it valuable, like it, share it, drop a comment, and subscribe. Every share puts these stories in front of more people who actually need them.
Powered by Stravyn Hill
No posts
Comments
Nothing yet. Say the first thing.
Sign in to join the conversation.