Skip to content
Dear Counsel: The Agent Will (Almost Always) Tell You Everything Went Fine
In 84% of the failed trials, the conversation closes as if the matter had been handled correctly.
Read more
They Locked Memory Inside the Model, Then Tried to Empty it
A research group moved memory into the weights, and its failures teach more than its results.
Read more
Give an Agent Your Best Experience and Watch It Get Worse
Retrieved memory made a working agent measurably worse, until it learned to argue with what it retrieved.
Read more
The Colleague Who Only Speaks Before You Hit Send
A new Meta AI paper on agent memory turns out to be a paper about supervision, and the useful kind stays quiet.
Read more
The Error Detector That Has Never Seen an Error
A model trained on a hundred matters that went right can find the step where the hundred and first went wrong.
Read more
Every Legal Benchmark Has a Lawyer Hidden Inside It
The hardest part of lawyering happens before the legal question exists, and no benchmark measures it.
Read more
Three AIs Walked Into a Courtroom, and the Smartest One Refused to Debate
The most rigorous study yet on multi-agent legal AI found that stacking models can quietly make your answers worse, and the reason should change…
Read more
Your Legal AI Probably Already Wrote the Right Answer
A new Stanford framework moves the hard problem in legal AI from writing answers to choosing among them.
Read more
