Memorilla: Latent Semantic Memory for LLMs
Introduces Memorilla, a memory system for LLMs that efficiently gathers and combines relevant information while keeping the context short.
Key points
- •Memory allows LLMs to utilize information outside their parameters, such as past conversations or agent actions.
- •For frozen LLMs, access to accumulated knowledge is limited to what the memory system places in the context window.
- •Existing systems face the challenge of gathering and combining relevant information for each query while keeping the context short.