Berkeley researchers teach AI agents to keep 'beliefs' instead of raw history for long tasks

A team of UC Berkeley researchers, including Jakob Björner, Aly Lidayan, Satvik Golechha, Kartik Goyal, and Alane Suhr, published a new framework called ABBEL on the BAIR blog on July 26, 2026, tackling a core limitation of AI agents that need to operate over long, multi-step tasks. As interactions grow, keeping a model's full history in context becomes impractical, and existing fixes like recursive summarization tend to lose important details, since the researchers found a summarizing model “never closes the gap” to a full-context model no matter how much it is trained. ABBEL instead has the agent maintain a natural-language belief state that it updates as new information arrives, and introduces a technique called belief grading that rewards the model for retaining whatever information it needs to reconstruct its most recent observation, rather than relying on hand-designed summarization rules. On CollabBench, a collaborative code-generation benchmark, ABBEL closed roughly half the performance gap to full-context models while needing well under half the peak memory, and it matched or beat full-context performance on a word-guessing task called Combination Lock. The authors argue that explicit belief states are also more interpretable and controllable than dense compressed context, potentially letting researchers directly inspect or edit what an agent believes rather than reverse-engineering it from a compressed vector. They frame the work as a step toward AI agents that can reliably operate over long horizons without runaway memory and compute costs, a prerequisite many labs see as necessary for trustworthy autonomous agents.