The Infra Pod Building the Future of AI with Long-term Memory (Chat with Charles from Letta)
Aug 11, 2025
Dive into the world of AI memory and discover its revolutionary potential. Charles from Letta discusses how long-term memory can enhance productivity tools and coding assistants. The concept of 'sleep time compute' is explored, highlighting continuous AI efficiency. Future trends suggest shared memory could outperform models in importance, shaping the way software interacts with users and improves collaboration. Tune in to learn about the early stages of AI memory technology and the challenges that lie ahead in creating more effective systems.
AI Snips
Chapters
Transcript
Episode notes
From Berkeley Lab To MemGPT
- Charles and his cofounders at Berkeley built MemGPT as an academic prototype that exposed the LLM memory problem.
- Their first prototype worked quickly and inspired Letta's mission to build long-term memory for agents.
Memory Is Token Management
- Charles reframes memory as tokens flowing in and out of the model, not just conversational state.
- He argues memory management is fundamentally token management across tools, reasoning, and context.
Memory Is Fundamental, Not Just A Tool
- Charles warns against treating memory as just another external tool like Google Drive.
- He says memory is too fundamental to AI to remain an optional tool called nondeterministically by clients.