Berkeley’s ABBEL trains agents to store graded belief states instead of full history

berkeley_ai · x · 2026-07-27

Berkeley presents ABBEL, a new way for agents to keep belief state as context grows

The Berkeley team argues that as task horizons get longer, LLM context windows cannot scale forever. Their new work, ABBEL, trains agents to maintain graded natural-language belief states instead of stuffing everything into the full history.

In other words, the agent learns to compress what it knows into structured beliefs, which is meant to make long-horizon behavior more manageable than simply carrying the entire interaction log forward.

Original post →

More from coding & agent

coding & agent channel →