Skip to content

Latest commit

 

History

History
134 lines (121 loc) · 13.9 KB

README.md

File metadata and controls

134 lines (121 loc) · 13.9 KB
Raw

🐨CoALA: Awesome Language Agents

Awesome License: MIT PR Welcome

teaser

A compilation of language agents using the Cognitive Architectures for Language Agents (🐨CoALA) framework.

@misc{sumers2023cognitive,
      title={Cognitive Architectures for Language Agents}, 
      author={Theodore Sumers and Shunyu Yao and Karthik Narasimhan and Thomas L. Griffiths},
      year={2023},
      eprint={2309.02427},
      archivePrefix={arXiv},
      primaryClass={cs.AI}
}

🐨CoALA Overview

CoALA neatly specifies a language agent starting with its action space, which has 2 parts:

  • External actions to interact with external environments (grounding)
  • Internal actions to interact with internal memories (reasoning, retrieval, learning)
    • A language agent has a short-term working memory and several (optional) long-term memories (episodic for experience, semantic for knowledge, procedural for code/LLM)
    • Reasoning = update working memory (with LLM)
    • Retrieval = read long-term memory
    • Learning = write long-term memory action_space

Then how does a language agent choose which action to take? Its actions are structured into decision making cycles, and each cycle has two stages:

  • Planning: The agent applies reasoning/retrieval actions to (iteratively) propose and evaluate actions, then select a learning/grounding action.
  • Execution: The selected learning/grounding action is executed to affect the internal memory or external world. decision_making

To understand more, read Section 4 of our paper.

Papers

Below is only a subset of papers scraped from CoALA.bib plus pulled requests, with potentially incorrect action space labels. Date is based on arxiv v1. They do not represent all language agent work, and we plan to add more work soon (pull requests welcome), and have labels for highly cited work.

(more to be added soon. pull request welcome.)

Resources

(more to be added soon. pull request welcome.)