Edge 415: Agents that Remember Actions with Procedural Memory
One of the most unique forms of memory in autonomous agents.
In this issue:
An overview of procedural memory in autonomos agents.
A review of Microsoft’s JARVIS-1, a memory-agumented LLM agents.
An introduction to the Zep framework for long term memory in LLMs.
💡 ML Concept of the Day: Procedural Memory in Autonomous Agents
To finalize this segment about memory in autonomous agents, today we will dive into procedural memory as a form of long-term memory. As it names indicates, procedural memory is related to procedures and actions an LLM agent will perform in a given environment.
Procedural memory has a strong parallel to human cognition as we learn to perform actions since we are babies. However, LLMs do not possess procedural memories in the conventional sense, such as knowing how to ride a bike or play basketball. However, they exhibit similar behaviors. For example, an LLM can recognize when to insert a line break at the end of each paragraph. Additionally, through fine-tuning, it is possible to teach an LLM to add specific phrases to its responses, like appending "Meow" to the end of every sentence.