Agent memory compression
Protected · $21,000Guarantee activeUS application · Filed Apr 13, 2026

Your attorney
The USPTO issued its first office action on Sep 24, 2026: two rejections, both over the same earlier publication. That is common and not final. I will draft our response and share it here well before the Dec 24, 2026 deadline.
What does this mean for us in plain terms?
Tao AI
The examiner thinks your main claim is too close to one earlier publication, and a second claim falls with it. Michael answers with arguments, narrower claims or both, and the application stays pending while he does.
The response is due Dec 24, 2026. Nothing is needed from you until he shares his draft.

Your attorney
My draft response narrows claim 1 to the utility scoring step, which that publication does not teach, and argues claim 9 on its own. Before I file, please confirm two technical statements:
1. Stored facts are scored again after every tool call.
2. An expanded fact is always fetched by its content hash.
Invention disclosure
Software output, not legal adviceInvention summary
A method for compressing an agent’s working memory into a bounded set of typed facts, scored by how useful they are downstream and expanded again on demand from a content addressed store, so long sessions keep their context without growing the prompt. Inventors: Ada Okafor, Tomás Reyes.
AI generated responses should not be treated as legal advice.
AI generated responses should not be treated as legal advice.