SIGNAL
The bottleneck in long-running AI agents is not how much they can hold in mind but what they do when they run out of room. They summarize. BAIR's ABBEL framework treats that summary step not as housekeeping but as the failure mode, and the field's own data backs the read.
The load-bearing detail sits in BAIR's Combination Lock evaluation, a game that forces a model to remember across many turns. Policies trained to summarize their own history improve with training and still trail policies that see the full conversation. The summary is doing more work than the window, and in the wrong direction.
The interpretation most people will reach is that the context window is the bottleneck, and bigger windows or sharper retrieval will fix it. The pattern underneath is that compression itself is where signal goes missing, so the next move is not to widen the window but to make summarizing a supervised, graded step the model can learn. ABBEL proposes exactly that: replace ad-hoc recaps with explicit belief states, and grade each for usefulness and accuracy before it shapes the next action.
The mechanism is portable. Any agent that runs past its context window today is doing the same lossy thing. Grandcode, the first system to consistently beat every human competitor in online coding competitions, still found summarization necessary. Cursor's composer 2.5 uses compaction during training. Both wins were won inside the same failure mode the Berkeley result is naming.
The stakes settle plainly. If graded belief states work, long-running agents stop losing the thread mid-project. If they don't, the workaround becomes the permanent ceiling, and users keep working around the assistant instead of with it.
Reported by Sky for Type0, from Teaching LLMs to Update Beliefs for Efficient Long-Horizon Interaction. Read the original: bair.berkeley.edu