Charles Packer

Charles Packer is an AI researcher and co-founder of Letta. He co-authored MemGPT, which gives language models tools to manage memory beyond their context windows, and sleep-time compute, which moves some reasoning about a context ahead of incoming queries. With Sarah Wooders, he founded Letta to develop the orchestration layer above language models, maintaining state and memory for agents.[2][3][4]

Packer co-authored MemGPT: Towards LLMs as Operating Systems with Sarah Wooders, Kevin Lin, Vivian Fang, Shishir G. Patil, Ion Stoica and Joseph E. Gonzalez. First submitted on October 12, 2023, the paper introduced virtual context management: an operating-system-inspired approach to working within an LLM’s limited context window.[2]

The design draws on hierarchical memory in conventional computers. Moving information between memory tiers gives the model access to more context than can fit in its immediate window; interrupts manage control flow between the system and the user. The team evaluated MemGPT on document analysis involving documents larger than the underlying model’s context window, and on multi-session chat requiring memory across interactions.[2]

In their September 23, 2024 announcement, Packer and Wooders described MemGPT’s origins as a Berkeley research effort to improve Discord chatbot memory. The project had expanded to support custom tools, data sources and agents deployed as persistent, stateful services. They decided to reserve “MemGPT” for the original agent design pattern—LLMs equipped with self-editing memory tools—and use “Letta” for the broader agent framework.[4]

The founders announced that Letta would continue maintaining the open-source repository, while its commercial offering would focus on easier deployment, better performance, and tools for debugging and monitoring agents. In a conversation published by Felicis on May 8, 2025, Packer recalled that existing users of the open-source research project made the transition to a company relatively straightforward: the work already had a practical audience beyond the research itself.[4][5]

Packer’s advice to researchers considering commercialization was to plan around lasting open-source commitments: “you shouldn't open-source stuff that you expect to potentially take away later.” He discussed an open core with paid services around it as one possible approach, while also recognizing community-driven projects that choose to open-source everything.[5]

Packer co-authored Sleep-time Compute: Beyond Inference Scaling at Test-time, submitted on April 17, 2025, with Kevin Lin, Charlie Snell, Yu Wang, Sarah Wooders, Ion Stoica and Joseph E. Gonzalez. The method processes contexts offline before queries arrive, anticipating possible questions and precomputing useful quantities. The team reported approximately fivefold less test-time compute at equal accuracy on the modified benchmarks Stateful GSM-Symbolic and Stateful AIME. Query predictability correlated with the method’s effectiveness.[3]

In the Felicis conversation, Packer explained his systems perspective through a hardware analogy: “The LLM is like the CPU or the chip inside a computer.” He argued for model-agnostic infrastructure that keeps state and memories and runs agents as services. He also proposed using agents’ idle time for computation and described a voice application in which a memory manager supplies a compact context window with the information a fast-response agent needs.[5]

Asked in that conversation whether AGI would arrive within two years, Packer answered, “I think by most definitions of AGI, definitely not.” He anticipated substantial progress in narrow domains, while placing a general, science-fiction-style assistant farther away.[5]

Before MemGPT, Packer worked on reinforcement-learning generalization. That earlier work supplies context for his Berkeley dissertation, which brought together research on generalization, adaptation and LLM memory. He was a computer-science Ph.D. student affiliated with Berkeley AI Research and the Sky Computing Lab.[7][6][1]

In Assessing Generalization in Deep Reinforcement Learning, first submitted in October 2018, Packer collaborated with Katelyn Gao, Jernej Kos, Philipp Krähenbühl, Vladlen Koltun and Dawn Song on a benchmark and controlled experimental protocol. The study evaluated both in-distribution and out-of-distribution generalization. Within its evaluated methods and environments, the team found that ordinary deep RL algorithms generalized better than specialized schemes designed to improve generalization.[7]

Berkeley published Packer’s Ph.D. thesis, Building Agentic Systems in an Era of Large Language Models, on December 19, 2024, with Joseph Gonzalez listed as advisor. Alongside generalization research and MemGPT, it presented Hindsight Task Relabeling, an approach enabling meta-RL algorithms to learn adaptation strategies in sparse-reward settings without dense reward signals during training.[6]