Most multi-agent LLM systems make agents talk in text, so every agent re-reads the whole history. What if they handed each other their internal memory, the KV-cache, instead? Four works from our lab give that memory four jobs: communicate, learn, score and check.

Does recursive self-improvement make AI agents forget? Often, yes. When LLM agents rewrite their own workflows, skills, weights or memory to master new tasks, they get worse at tasks they had already mastered. We call it capability erosion under self-evolution.

Math is not just about problem solving. Last week, OpenAI's release 722 manuscripts is a real milestone, but problem solving is only one part of mathematics. We tested an understanding: can a language model keep track of what a theorem assumes when it moves the theorem into a neighbouring field? Across seven frontier LLMs, the answer was mostly no. When they restated a theorem in a more general setting, 60.6% of their rewrites silently dropped a hidden hypothesis, and not one erred in the opposite direction.