Do AI Agents Dream of Memory?

During sleep, the human brain sorts through memories, consolidating important ones and discarding others. This biological process inspires the idea that AI could perform similar memory consolidation.

Introducing such capability, Bilt, a company offering local deals to renters, has deployed millions of AI agents to emulate this natural process. These agents, powered by Letta, leverage "sleeptime compute" to determine which data should be stored long-term for efficient recall when needed.

Andrew Fitz, an AI engineer at Bilt, explains that a single update to a memory block can transform the operations of thousands of agents, highlighting the importance of granular control over AI agent contexts.

Large language models are limited in their recall abilities, reliant on the context window into which information must be explicitly pasted for retrieval. They often struggle with managing large volumes of context, unlike the human brain that stores information for later recall.

Letta’s technology aims to mirror biological memory processes, with Charles Packer, Letta’s CEO, affirming that while human brains continuously improve, language models accumulate noise and can derail without context management.

Packer and co-founder Sarah Wooders had previously worked on MemGPT, an open-source LLM project. With Letta, they enhance these concepts to improve AI agents' memory handling, as seen in their collaboration with Bilt.

Memory development remains a key area in AI advancement, with companies like LangChain also contributing innovations. LangChain enhances AI memory use through context engineering, offering diverse memory storage options tailored for AI agents.

OpenAI has also announced improvements in ChatGPT’s memory, aiming to provide more personalized user experiences. While mechanisms remain undisclosed, these developments signify a shift toward smarter and less error-prone AI.

Clem Delangue, CEO of Hugging Face, stresses the need for open memory systems, suggesting transparency can push AI technology forward.

Interestingly, Packer emphasizes the importance of forgetting, proposing that AI should erase specific memories upon user requests, reflecting a facet of human memory dynamics.

These advancements echo the speculative concept from Philip K. Dick's novel, "Do Androids Dream of Electric Sheep?"—inspiring reflections on the evolving sophistication of AI memories akin to fictional androids.